385'8(81,9(56,7< *5$'8$7(6&+22/ 7khvlv...
Post on 19-Mar-2018
215 Views
Preview:
TRANSCRIPT
01 14
PURDUE UNIVERSITY GRADUATE SCHOOL
Thesis/Dissertation Acceptance
Thesis/Dissertation Agreement.Publication Delay, and Certification/Disclaimer (Graduate School Form 32)adheres to the provisions of
Department
Christian Eduardo Silva Salazar
Development, Numerical Demonstration, and Experimental Verification of a Method for ModelUpdating of Boundary Conditions.
Master of Science in Mechanical Engineering
Shirley J. Dyke
Arun Prakash
Jeffrey F. Rhoads
Shirley J. Dyke
David Anderson 07/29/2014
DEVELOPMENT, NUMERICAL DEMONSTRATION AND EXPERIMENTAL
VERIFICATION OF A METHOD FOR MODEL UPDATING OF BOUNDARY
CONDITIONS
A Thesis
Submitted to the Faculty
of
Purdue University
by
Christian E. Silva
In Partial Fulfillment of the
Requirements for the Degree
of
Master of Science in Mechanical Engineering
August 2014
Purdue University
West Lafayette, Indiana
ii
To God.
To Ximena, Tomas, Ana and Marıa.
To my parents and sisters.
iii
ACKNOWLEDGMENTS
Although trying to acknowledge every person that somehow has contributed to
this work would be impossible, I would like to dedicate special recognition to my
advisor, Dr. Shirley Dyke whose help and support beyond the academic realm have
been cornerstone in conceiving, developing and finishing this thesis. I would also like
to thank the faculty members of my committee who were kind enough to accept this
appointment, to read and comment my thesis and to evaluate my work. A special
note will be dedicated to Dr. Juan Caicedo at University of South Carolina for his
contribution with the MATLAB Finite Element Toolbox used in this work, and to Dr.
Arun Prakash, for his generous sharing of information for the finite element analysis.
I owe a great deal of appreciation and gratitude towards my colleagues in the IISL
group; without their selfless suggestions, ideas and support, this task would have been
a much heavier burden.
I will also include the SENESCYT, my home Government funding agency as an
important part of this achievement. I am convinced that their commitment and sup-
port towards fellow Ecuadorian graduate students all over the globe will contribute
greatly in the growth of a better Ecuador for future generations.
Special thanks will be given to the National Science Foundation, grant number
NSF-CNS-1035748, the Luna-Army SBIR funding for supporting me during my stud-
ies, and the College of Engineering and the Mechanical Engineering Department at
Purdue University for providing the funding for the IISL.
Finally, I will thank my family who agreed to leave everything they knew as life to em-
bark in this adventure called Graduate School. It sure has been a learning experience
for the five of us.
iv
TABLE OF CONTENTS
Page
LIST OF TABLES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi
LIST OF FIGURES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii
SYMBOLS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . x
ABBREVIATIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi
ABSTRACT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xii
1. INTRODUCTION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31.2 Purpose of Project . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
Part I: DEVELOPMENT
2. LITERATURE REVIEW . . . . . . . . . . . . . . . . . . . . . . . . . . 82.1 The Vibration of Euler-Bernoulli Beams with Different Boundary Con-
ditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82.2 Modal Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2.1 Frequency-Domain Methods for Modal Analysis . . . . . . . 142.2.2 Time-Domain Methods for Modal analysis . . . . . . . . . . 16
2.3 Modal Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3. THEORETICAL BACKGROUND . . . . . . . . . . . . . . . . . . . . . 213.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213.2 The Fourier Transform . . . . . . . . . . . . . . . . . . . . . . . . . 213.3 Frequency Response Function . . . . . . . . . . . . . . . . . . . . . 223.4 The Peak Amplitude Method . . . . . . . . . . . . . . . . . . . . . 24
4. A CORRELATION METHOD . . . . . . . . . . . . . . . . . . . . . . . 274.1 Correlation based on the Modal Assurance Criterion . . . . . . . . . 274.2 Resolution Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
Part II: NUMERICAL VALIDATION
5. MATHEMATICAL MODELS . . . . . . . . . . . . . . . . . . . . . . . . 335.1 Analytical Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
5.1.1 Simply-Supported Beam Model . . . . . . . . . . . . . . . . 355.1.2 Beam with a Torsional Spring Support at Each End . . . . . 385.1.3 Beam with Rotational Masses on Both Ends . . . . . . . . . 40
v
Page
5.2 Finite Element Models . . . . . . . . . . . . . . . . . . . . . . . . . 435.2.1 Simply-Supported Beam Model . . . . . . . . . . . . . . . . 435.2.2 Beam With a Torsional Spring Support at Each End . . . . 475.2.3 Beam With Rotational Masses on Both Ends . . . . . . . . . 505.2.4 Finite Element Model Details . . . . . . . . . . . . . . . . . 52
6. NUMERICAL VERIFICATION . . . . . . . . . . . . . . . . . . . . . . . 596.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 596.2 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 596.3 Model Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 616.4 Correlation Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 62
Part III: EXPERIMENTAL VALIDATION
7. EXPERIMENTAL SETUP . . . . . . . . . . . . . . . . . . . . . . . . . . 677.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 677.2 Conceptualization & Design Parameters . . . . . . . . . . . . . . . 677.3 Research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 687.4 Design and Construction . . . . . . . . . . . . . . . . . . . . . . . . 697.5 Assembly . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
8. STRUCTURAL IDENTIFICATION AND CORRELATION OF CASE STUDY 748.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 748.2 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . 758.3 Algorithm for Modal Identification Fast Fourier Transform . . . . . 78
8.3.1 Modal Analysis Results of Experimental Setup . . . . . . . . 828.4 Correlation to a Simply-Supported Beam . . . . . . . . . . . . . . . 83
8.4.1 Graphic Results for the Simply-Supported Case . . . . . . . 858.4.2 Correlation Results for the Simply-Supported Case . . . . . 858.4.3 Correlation to a Clamped-Clamped Beam . . . . . . . . . . 888.4.4 Correlation to a Beam with Rotational Masses at Both Ends 91
8.5 Correlation Algorithm to Determine Model Parameters . . . . . . . 948.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 968.7 Implementation of the Methodology . . . . . . . . . . . . . . . . . . 98
9. CONCLUSIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
10.RECOMMENDATIONS AND FUTURE DIRECTIONS . . . . . . . . . 102
LIST OF REFERENCES . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
VITA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
vi
LIST OF TABLES
Table Page
5.1 FEM frequency comparison (in Hertz). . . . . . . . . . . . . . . . . . . 47
5.2 FEM frequency comparison for a system with rotational springs on bothends (in Hertz). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
5.3 FEM frequency comparison for a system with rotational masses on bothends (in Hertz). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
8.1 Experimental configurations. . . . . . . . . . . . . . . . . . . . . . . . . 75
8.2 Final frequency comparison (in Hertz). . . . . . . . . . . . . . . . . . . 96
vii
LIST OF FIGURES
Figure Page
1.1 Model updating and parallel techniques. . . . . . . . . . . . . . . . . . 2
2.1 Frequency-domain methods, (from [50]). . . . . . . . . . . . . . . . . . 12
2.2 Time-domain methods, (from [50]). . . . . . . . . . . . . . . . . . . . . 13
2.3 Modes manually picked from a frequency response function. . . . . . . 15
3.1 Single DOF system. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.2 The half-power method. . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.3 Estimation of the first three mode shapes of a cantilever beam with threesensor locations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.1 An example of a 5 × 5 correlation matrix. . . . . . . . . . . . . . . . . 30
4.2 Resolution sensitivity of the proposed method. . . . . . . . . . . . . . . 32
5.1 Classic boundary conditions. . . . . . . . . . . . . . . . . . . . . . . . . 34
5.2 First four modes of a simply-supported beam. . . . . . . . . . . . . . . 38
5.3 Beam on elastic supports. . . . . . . . . . . . . . . . . . . . . . . . . . 39
5.4 Limiting mode shapes for vibrations of a beam on elastic supports (solidlines: simply-supported; dashed lines: clamped-clamped). . . . . . . . . 40
5.5 Simply-supported beam with rotational masses on both ends. . . . . . 41
5.6 First four mode shapes of beam with rotational masses on both ends. . 42
5.7 FEM node layout. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
5.8 Simply-supported case mode shapes generated by FEM. . . . . . . . . 47
5.9 Layout of elements generated by the FEM. . . . . . . . . . . . . . . . . 49
5.10 Elastic support case mode shapes generated by the FEM. . . . . . . . . 50
5.11 Modeshapes of a simply-supported beam with rotational masses on bothends, generated by the FEM. . . . . . . . . . . . . . . . . . . . . . . . 52
5.12 Beam element with element coordinates. . . . . . . . . . . . . . . . . . 53
viii
Figure Page
5.13 Detail of the FEM rigid links. . . . . . . . . . . . . . . . . . . . . . . . 57
5.14 Boundary condition definition example. . . . . . . . . . . . . . . . . . . 58
5.15 Rotational mass definition example. . . . . . . . . . . . . . . . . . . . . 58
6.1 Modal vector comparison. . . . . . . . . . . . . . . . . . . . . . . . . . 62
6.2 Case study. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
6.3 Correlation comparison of spring-supported and rotational-mass parame-ter combinations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
6.4 MAC-based correlation results. . . . . . . . . . . . . . . . . . . . . . . 65
6.5 MATLAB command window results. . . . . . . . . . . . . . . . . . . . 66
7.1 Scaled truss bridge for SHM. (Courtesy of Dr. Jennifer A. Rice, Universityof Florida at Gainesville). . . . . . . . . . . . . . . . . . . . . . . . . . 69
7.2 Scaled bridge. Conceptual design. . . . . . . . . . . . . . . . . . . . . . 70
7.3 Detail of mechanism for modifying the boundary conditions: bolted =fixed; unbolted = pinned. . . . . . . . . . . . . . . . . . . . . . . . . . 72
7.4 Aerial view of the structure. . . . . . . . . . . . . . . . . . . . . . . . . 73
8.1 Experimental configurations. . . . . . . . . . . . . . . . . . . . . . . . . 76
8.2 Hammer response curves (from PCB Piezotronics). . . . . . . . . . . . 78
8.3 Experimental input/output signals. . . . . . . . . . . . . . . . . . . . . 79
8.4 Hammer hit peak choosing. . . . . . . . . . . . . . . . . . . . . . . . . 80
8.5 Number of peaks within an imput window. . . . . . . . . . . . . . . . . 81
8.6 FFT code outputs. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
8.7 First mode shape vector along with its graphical representation. SS model. 84
8.8 Second mode shape vector along with its graphical representation. SSmodel. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
8.9 Third mode shape vector along with its graphical representation. SSmodel. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
8.10 Vector of two stacked mode shapes along with its graphical representation.SS model. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
8.11 Simply-supported vs. experimental mode shapes. . . . . . . . . . . . . 88
8.12 Correlation results. Simply-supported model. . . . . . . . . . . . . . . 88
ix
Figure Page
8.13 Modeshapes of a clamped-clamped beam. . . . . . . . . . . . . . . . . 89
8.14 Clamped-clamped vs. experimental mode shapes. . . . . . . . . . . . . 90
8.15 Correlation results. Clamped-clamped model. . . . . . . . . . . . . . . 90
8.16 First three mode shapes of a beam with asymmetrical rotational masses. 92
8.17 Rotational mass vs. experimental mode shapes. . . . . . . . . . . . . . 93
8.18 Correlation results. Rotational mass model. . . . . . . . . . . . . . . . 93
8.19 Correlation comparison of nine parameter combinations. . . . . . . . . 95
8.20 MAC-based correlation results. . . . . . . . . . . . . . . . . . . . . . . 97
8.21 Correlation method. Implementation . . . . . . . . . . . . . . . . . . . 99
x
SYMBOLS
m mass
x position in the horizontal axis
v velocity
a acceleration
ω natural frequency
φ eigenvector (mode shape)
ψ mode shape
ζ damping ratio
E modulus of elasticity
Ix second moment of area around x
Iy second moment of area around y
Iθ rotational inertia
G shear modulus
kt linear spring constant
kθ rotational spring constant
β eigenvalue (natural frequency)
A area
w deflection as a function of time and position
ke finite element stiffness matrix
me finite element mass matrix
A FEM global matrix assembly
xi
ABBREVIATIONS
kg kilograms
m meters
rad radians
Hz hertz
N Newtons
FRF frequency response function
IFRF inverse frequency response function
TF transfer function
SS state space
FT Fourier transform
FFT fast Fourier transform
IFFT inverse fast Fourier transform
DFT Discrete Fourier transform
ERA eigensystem realization algorithm
Re real
Im imaginary
xii
ABSTRACT
Silva, Christian E. M.S.M.E, Purdue University, August 2014. Development, Numer-ical Demonstration and Experimental Verification of a Method for Model Updatingof Boundary Conditions. Major Professor: Shirley J. Dyke, School of MechanicalEngineering.
The study of vibrations in beams has been largely addressed by authors and
researchers. However, relatively few researchers have considered the case of unknown
boundary conditions, as usually it is reasonable to assume the classical cases such
as simply supported, clamped or free. Indeed, there are a wide variety of boundary-
condition configurations, each one representing a whole different problem with its
own modal characteristics.
A method for updating experimental beam models to specifically address the is-
sue of unknown boundary conditions is proposed. This methodology takes advantage
of vector comparison techniques such as the modal assurance criterion and the dot
product to determine the degree of linear relationship between two mode shapes sys-
tematically and iteratively until an acceptable parametric match is found. This thesis
includes the phases of development, numerical demonstration and experimental veri-
fication. In the section devoted to development, a detailed explanation of the method
is given; the numerical demonstration section is intended to demonstrate the capabil-
ities of the method using mathematical models only; and finally, in the experimental
verification section, a case study example is developed using real experiment data.
At the end of this thesis, a generalized procedure is described so the method can be
applied to beam-behaving structures and ultimately any engineering model in which
boundary conditions have an important role.
1
1. INTRODUCTION
A hundred years ago, medicine relied mostly on the physical build of an individual’s
body and had almost no influence on one’s life expectancy. However, nowadays
medicine is not only a means of monitoring a person’s health, but of extending a life
through appropriate corrective procedures. Similarly, structures are viewed now in a
different light than in the past. The life of structures can be not only monitored but
also extended, if adequate monitoring techniques are used to determine a structure’s
condition. “The process of implementing a damage detection and characterization
strategy for engineering structures is referred to as Structural Health Monitoring
(SHM)” [1]. SHM utilizes techniques and results obtained from two closely related
fields that work side-by-side: modal analysis and finite element model updating.
“Modal analysis is the field of measuring and analyzing the dynamic response of
structures when excited” [2]; and ”finite element model updating is the process of
ensuring that finite element analysis results in models that better reflect the measured
data than the initial models” [3].
Lately, the process of building a structure, whether it is static (a building) or
dynamic (a machine part), has shifted from being a one-time action that ended at
the service start-date of the structure and relied mainly in the durability of the
selected materials, to a more interactive process in which the initial design is one
of many stages in the structure’s service lifetime. Indeed, nowadays a great deal of
information from the structure can be obtained which is further used for improving
or correcting its behavior.
To properly predict the behavior of a structure, a suitable model must be se-
lected. This step requires experience and some understanding of the structure’s be-
havior. Even though many models have already been developed by engineers and
researchers, no model can be picked off-the-shelf for a determined problem. Every
2
structure, no matter how similar it is to a previous one, has a different behavior due
to uncertainties that are present from many sources: material non-homogeneities,
environmental surroundings, usage, assembly discrepancies and of course, boundary
conditions. Therefore, the selected model must be corrected so it can properly predict
the system’s behavior. This process can be a dynamic one, that is to say that the
parameters of the model can be updated many times during a structure’s lifespan.
But to determine how much a model has to be modified or which parameter should
be updated, after having some knowledge of the behavior of the structure, some sort
of identification process has to be performed. Figure 1.1 shows a schematic diagram
of the described processes.
Figure 1.1. Model updating and parallel techniques.
Once a model has been selected, and the structural behavior has led to a model
updating, such a structure can begin its service life and its response can start to
be monitored so that its behavior is continuously compared to the predicted by the
updated model. This process is called structural health monitoring, and it is among
the most important subjects in civil, mechanical and aerospace engineering and has
3
received increasing attention from researchers and engineers in recent years. There
are two different types of methodologies for health monitoring; namely, localized
and global techniques [4]. Localized techniques are based on specific location analy-
ses, such as observations and non-destructive tests performed in key locations of the
structure; whereas global techniques analyze the structure as a whole, studying, for
instance, its frequencies or mode shapes for determining its condition.
Modern sensor and communication technologies, as well as computational capa-
bilities, have made it possible for an engineer to know the conditions of a structure
in real-time and, therefore, since every structure is changing over time, to know with
high accuracy any deviations from its intended mechanical and dynamic properties
at the very moment of appearance of such deviations. The retrieved information can
be used to schedule adequate and cost-effective maintenance works, to plan and exe-
cute any corrective modifications to the physical structure, and even to estimate the
remaining life with a fair amount of accuracy.
Structural dynamics or modal analysis is the traditional method for obtaining
dynamic properties of a given system or structure, and constitutes an important step
for model updating. Several modal identification techniques have been developed,
some of which will be used in the present thesis to implement the proposed method.
1.1 Overview
This project addresses the problem of a structure whose model has been partially
defined. Preliminarily, a simply supported beam model has been selected, but there
exist significant deviations between the model’s modal information the experimental
results. The idea of changing a parameter of the structure for model updating though
promising, does not provide a realistic technique because of the large deviations noted
above. Therefore, an alternative technique for determining a better model has to be
developed beforehand.
4
The goal of this project is to investigate a way to provide a model updating tech-
nique that takes into account the problem of boundary conditions instead of structural
parameters, and how a change in these boundary conditions will affect the behavior
of the structure. The steps to achieve the proposed goal include the development
of an adequate methodology for comparing modal parameters in a vectorial fashion;
the verification of such a methodology with information from mathematical models
only to obtain perfect matches; and, after confirming the strength of the method, the
extension of the procedure to experimental data to try to obtain a mathematically
close relationship between the experimental and analytical models.
The success of the present project is based in many aspects, such as the avail-
ability of analytical models already developed which only have to be compared to
that of the experiment, and the vectorial form of modal parameters, either in time
or frequency domain, which can be straightforwardly compared using an appropriate
computer algorithm until the best match is found. Some of the assumptions made
for this project are the models of the structure. A group of Euler-Bernoulli models
of bending vibrations in beams are considered for the present thesis. Moreover, there
are around fifty analytical cases that can be included in the comparison algorithm
which constitutes a fair large analytical ‘model space’ to choose form. If each one
of these models is tested with different parameters such as rotational inertias, linear
or rotational spring stiffnesses, lumped masses, or combinations of these, the “model
space” can reach several hundreds of possibilities. As any proposed technique, the
methodology of this thesis is not exempt from risks and obstacles, such as the travel
from the continuous domain to the discrete domain in the experimental setup. A
beam is to be tested by the use of a finite number of sensors which will not repli-
cate the continuous behavior exactly, leaving a space for uncertainties and unknown
parameters. Nevertheless, the proposed method is not a model validation technique,
but more of a model updating technique which instead of modifying parameters of
a set model, shifts to a different model (beam model + boundary conditions) within
5
a reasonable error margin. For this study, four beam models have been chosen: two
classical and two non-classical, which are listed below:
• Simply-supported: both ends are pinned with translation constraints in x, y and
z.
• Clamped-clamped: both ends are fixed with translation and rotation constraints
in x, y and z.
• Partial elastic supports: ends are pinned and have rotational springs of constant
kθi on both ends (i = 1, 2).
• Mixed inertial supports: pinned ends with inertial rotational masses of constant
Iθi on both ends (i = 1, 2).
The thesis is divided in three sections: Development of the Method; Numerical
Validation and Experimental Verification
The first section includes Chapters 2, 3 and 4. Chapter 2 is a short review of
related literature; Chapter 3 includes some of the theoretical background on which the
thesis is based; and the basics of the method are explained in Chapter 4. The second
section is composed of Chapters 5 and 6. In Chapter 5, three mathematical models
are developed using both analytical and finite element methods for understanding
the mechanics of boundary condition changes in beams. To ensure the functionality
of the method, Chapter 6 shows how the method works with data collected through
simulations from analytical models. The third section includes Chapters 7 and 8. In
Chapter 7 a short description of the structure designed for this study is developed,
whereas Chapter 8 deals with the application of the method with an experimental
example. Final chapters devoted to conclusions and future directions of research are
included outside of the mentioned sections.
6
1.2 Purpose of Project
The specific purpose of this project is to develop a method for model updating
which iteratively compares the response of a structure with the responses of a group of
selected structures whose models have been preselected. The aim of such comparison
is to detect the highest correlation between the case study model and any of the
selected models.
The proposed method is a very interesting approach because it can be applied
using a “solution space” or “model space”, which can be of any finite size. All of
the elements of this “model space” are compared to the specimen model (case study)
one by one to construct a correlation vector or matrix which, if further analyzed,
determines which model is more likely to be generating such a response. At this
point, it is important to make a clarification: although the proposed method tries to
correlate so-called ‘models’, this is indeed a model updating methodology a bit away
from traditional model updating techniques inasmuch as the parameters that are tried
to be correlated are boundary condition parameters, instead of a structural physical
parameters. Consider for example a model with the mass, stiffness and damping
matrices, M,C,K respectively are modified accordingly for model updating using a
least squares approach so that the model matches some experimental behavior; the
proposed technique will not change any beam structural parameters directly (such as
M,C or K), but some boundary-condition parameters (such as a spring constants or
rotational mass inertias or lumped masses at fixed points) to obtain the same desired
behavior. This constitutes a novel idea since boundary conditions are more likely to
be the reasons for a model deviation instead of intrinsic material structural matrices.
For the case of this thesis, only four models were studied and used as reference, two
of which had variable parameters to construct a fair large model space for comparison,
but in a real life case, many more can be used. Consider for example the case
of a beam with a concentrated mass at midspan and a distributed load along its
length: The first most intuitive model that can be considered is a simply-supported
7
beam with a lumped mass at midspan, but if after performing modal testing on this
hypothetical structure an engineer realizes that its parameters don’t match those of
his first selected model, he has to repeat the modeling process over and over again until
a good approximation appears, resulting in a very tedious time consuming process.
On the other hand, by using the proposed method, it is only matter of running a set of
correlation analysis modifying the combinations lumped mass/concentrated load until
a high enough index is obtained which is an indicative of the closest mathematical
match.
PART I: DEVELOPMENT
8
2. LITERATURE REVIEW
2.1 The Vibration of Euler-Bernoulli Beams with Different Boundary
Conditions
The vibration of beams with classical boundary conditions is a topic that has
been widely covered by various authors, both in books and scientific articles. How-
ever, when non-classical boundary conditions are present, the available bibliography
becomes rather limited. Analytical solutions of the beam problem with simple bound-
ary conditions such as pinned, fixed and sliding configurations were addressed in books
by Meirovitch [5], Karnovsky [6], Gorman [7] and Blevins [8], just to cite a few.
Although general boundary conditions such as combinations of pinned, fixed and
sliding ends appear to be simple cases, these are nothing but special cases of more
complex analyses involving elastic boundary conditions in the form of springs, both
linear and rotational, located either at the ends of a beam, or at intermediate points.
For instance, a simply-supported beam is a special case of a beam resting on linear
springs of infinite stiffness which provide the translational constraint in the x, y
and z directions, and rotational springs both of zero stiffness which provide the free
rotation around the y axis and rotational springs of infinite angular stiffness which
provide the rotational constraints around the x and y directions. Several studies
have been published on these types of constrained beams. Hibbeler [9] considered the
vibration of a beam limited at both ends with rotational springs of different stiffnesses,
although his results were later corrected by himself [10], and documented by several
other authors like Goel [11] and Rao et al. [12]; Chun [13] derived the eigenvalues
and eigenvectors of a vibrating beam with one end constrained by rotational and
translational springs, and the other end free, whereas Lv et al. [14] and Kang &
9
Kim [15] studied general boundary condition cases. For the case of a mass-beam
system, i.e., a beam with a mass at a certain distance from the origin, Goel [11]
developed both analytical equations and eigenfrequency tables for different spring
constant rates. Many other authors developed related works involving mass-beam
systems; some examples are the publications by Register [16], Yen [17], Goel [11] and
Laura et al. [18]. Maurizi, along with other researchers, developed an extensive work
on vibrations of beams publishing several articles. Maurizi studied a case where a
beam is constrained with a rotational spring on one end, and a translational spring
on the other; i.e., mixed boundary conditions on the same beam [19]; other case was a
beam with elastic constraints on both ends subjected to free or forced vibrations [20];
and he even did a review of many boundary condition cases published by previous
authors [21]. All these cases are more realistic scenarios for boundary conditions to
which structures are subjected on a daily basis. One good example of a problem where
boundary conditions are not ideal is an emergency deployable bridge structure which
is ‘placed’ over the connecting points without any soil or ramp preparation. In this
case, no information about boundary conditions is available despite the fact that these
boundary conditions are precisely the parameters which give the system its natural
characteristics. Moreover, these boundary conditions will most likely change on a
regular basis due to usage, weather conditions and material degradation. Another
example is a beam connected to rigid supports by pins (simply-supported) in which
the pin-hole assembly deviates from its design leading into unknown forcing moments.
As the reader can imagine, boundary conditions are a very delicate problem that has
to be addressed with the maximum of accuracy in a real-life structure.
The previous literature review dealt with general analytical and numerical meth-
ods for solving the governing equations along with the boundary condition equations.
However, there is also a great deal of research regarding alternative methods for solv-
ing the problem of vibrations of beams with both classical and elastic constrained
ends. One of these methods is Fourier series. A general methodology for structural
10
vibration problems using Fourier series was published by Greif and Mittendorf in
1976 [22]. Li [23] solved the problem of vibration of a generally supported beam by
combining Fourier series and an auxiliary polynomial function. Wang [24] solved this
problem in a similar way, but also using Stoke’s transformations for the derivatives
in the boundary conditions; and more recently, Yayli et al. [25] and [26] extended
this approach for elastically-constrained ends and elastic foundation, which are not
covered in the present thesis.
Other alternate methodologies, for solving the vibration problem of an Euler-Bernoulli
beam subjected to various boundary condition combinations were used by Lai et
al., [27] and Mao [28]. They used an Adomian decomposition to transform the gov-
erning differential equation of a beam into a recursive algebraic equation which con-
verts the boundary conditions into simple algebraic equations suitable for symbolic
computation. Although the mathematical effort to derive the frequency equations is
not very straightforward, once these have been obtained, this method becomes very
simple when a different boundary condition case is to be analyzed, making the overall
process easier and less time consuming.
So far, the cited bibliography referred only to continuous beams. Many publica-
tions are available for the problem of vibrations of discontinuous beams. Failla and
Santini [29] proposed a solution method for beams with internal translational and
rotational springs whose work was based primarily in Billelo and Bergman’s study of
crack modeling using internal rotational and translational springs [30]. Other good
examples of this approximation for cracks on beams with springs are the studies by
Raffo and Carrizo which introduced the solution of an inverse problem for beams and
frames under presence of cracks [31]; Wang & Qiao, who proposed a Laplace trans-
form approach for the solution of the governing equations [32]; and Ratazzi et al.,
who extended their work for frames with internal hinges [33]. A variation of discon-
tinuous beams is the case of multispan beams. Several authors covered special cases
11
of multi-span beams with intermediate supports. Again, Karnovsky, [6], Gorman [7]
and Blevins [8] devoted separate chapters in their books for single, double, triple,
quadruple and multi-span beams with combined boundary conditions and even with
inter-span spring connections.
Some cases of cantilever beams are covered but as particular cases of more general
ones. A rather large amount of publications have been made regarding to cantilever
beams as well. Also, this literature deals with problems in which the authors use
a combination of analytical and numerical methods for solving the equations. For
example, in most cases, the governing partial differential equation of the beam is
solved using the principle of separation of variables. However, once the equation of
frequency is obtained, this is usually a transcendental equation whose solution cannot
be found analytically and a numerical method like bisection or Newton-Raphson has
to be used for root-finding. Moreover, approximate methods have been used also to
solve the complete problem, in the form of finite element analysis. Meirovitch [5]
devotes a chapter of his book for describing the finite element method as a means
of modal analysis using the Rayleigh-Ritz method and both linear and higher degree
interpolation functions. Other good sources for theoretical background and practical
implementation of finite element analysis in vibration problems are the books by
Courant [34], Huebner [35] and Zienkiewicz & Taylor [36].
2.2 Modal Analysis
Modal analysis is a foundational section of system identification. “It is a process of
analysis of the vibratory behavior of structures and systems by means of theoretical
or technical techniques. The ultimate goal is the determination of a structure or
system modal parameters for further constructing a mathematical model of such a
system,” [37] with the objective of “determining, improving and optimizing dynamic
characteristics of engineering systems and structures” [38]. It has spread so widely
12
that numerous methods for modal identification have been developed and published
by authors throughout the last four decades [38] [39].
Brown & Allemang [40] considered that modal analysis is founded in two main
historical scientific breakthroughs: the decomposition of solar spectrum in its color
components made by Newton, and the subsequent decomposition of arbitrary func-
tions into its simple harmonic components made by Fourier. However, it wasn’t until
the mid twentieth century where modal analysis started to gain academic and in-
dustrial interest due to the necessity of understanding the vibrations on airplanes,
machine tools and vehicles. The introduction of the fast Fourier transform method
by Cooly & Tukey [44] could be considered a turning point of the fast development
and increasing popularity of modal analysis as a previous step for various applica-
tions, such as damage detection, structural health monitoring, structural control and
structural design. It is not the intention of the present thesis to do a historical review
of modal analysis, but to cite just a few relevant findings in the area. As a conse-
quence of the popularity of the field, many books have been published like the ones
by Mendez Maia [41], Ewins [42] and Fu & He [43].
Figure 2.1. Frequency-domain methods, (from [50]).
13
There is a wide list of modal analysis techniques developed recently and a brief
description of these will be made herein. Nevertheless, it will be helpful to provide the
big picture of modal identification methods. The experimental direction frequently
used by researchers is to collect acceleration, velocity or position data from a sensor
(output), when an excitation (input) has been imposed to the structure to obtain, af-
ter proper analysis, an input/output relationship called transfer function (TF) which
contains the modal characteristics of the structure (frequencies, mode shapes and
damping ratios) or the mechanical characteristics of the system (mass and stiffness).
Excitations can be as varied as ambient (traffic, wind, plant noise), impact (instru-
mented hammers or exciters), known functions (harmonic, step, earthquake history),
etc. Modal analysis can be performed both in the time-domain and in the frequency-
domain. These two main divisions can also be sub-divided into direct and indirect
methods; Figures 2.1 and 2.2 depict an overall idea of most of the available methods.
Frequency domain methods use the frequency response function (FRF) information
and time-domain methods use finite element analysis or the inverse FRF which is
nothing but the inverse fast Fourier (IFFT) transform of the FRF.
Figure 2.2. Time-domain methods, (from [50]).
14
2.2.1 Frequency-Domain Methods for Modal Analysis
Frequency-domain modal analysis uses a mathematical model that represents the
frequency response function (FRF) data from measurements. An analytical expres-
sion of the FRF or commonly known as transfer function, as a partial fraction expan-
sion can be given by:
H(ω) =N∑r=1
Ak
iω − βk+
A∗kiω − β∗k
(2.1)
where Ak is the modal constant matrix with components rAij related to components
φik, φjk of eigenvector k through the formula:
rAija = qkφikφjk (2.2)
and qk is a modal participation factor. Index k indicates the k-th mode and the poles
in the denominator can be obtained by solving the characteristic equation
βk, β∗k = −ωkζk ± i
(ωk
√1− ζ2k
). (2.3)
This implies that most frequency-domain methods take advantage of the infor-
mation given by a transfer function to derive the modal characteristics of a linear
time-invariant system. Combinations of inputs, outputs and locations lead to a set
of FRF’s. The most commonly used frequency-domain techniques for modal identifi-
cation are peak-picking, circle fit, inverse FRF, least squares and Dobson’s method.
The simplest frequency-domain method for identifying modal parameters is the
peak picking method, whose idea is to manually pick the peaks of the FRF which
correspond to the resonance frequencies. The damping ratios can be estimated by
evaluating the sharpness of such points; and the mode shapes, from the relation
between peak amplitudes in the structure [45]. This method was proposed by Bishop
& Gladwell [46]. Unfortunately, this approach works fine for well separated and
15
clear modes whereas for complex structures that mix lateral with torsional modes
may not be very efficient. Nevertheless, it is very popular due to its simplicity and
implementation time. The present work utilizes also this method for mode verification
and elimination of torsional modes by observations made in the imaginary part of the
transfer function. This method is shown in Fig. 2.3.
0 50 100 150−150
−140
−130
−120
−110
−100
−90
Frequency, Hz
Mag
nigu
de, d
B
X: 16.98Y: −91.76 X: 46.6
Y: −100
X: 91.7Y: −105.8
Figure 2.3. Modes manually picked from a frequency response function.
Another important method developed by Kennedy [47] is the circle fit, which uses
the relationship between real and imaginary part of a transfer function around each
natural frequency to identify modal parameters of the system. This method’s basic
idea is that the receptance FRF α(ω) traces a perfect circle on the Nyquist plane,
and where h is a measure of structural damping.
[Re(α)]2 +
[Im(α) +
1
2h
]2=
(1
2h
)2
. (2.4)
A very interesting approach for modal estimation is derived from the properties
of the inverse of the frequency response function (IFRF) which are nothing but the
Bode plots presented by Dobson [48]. By plotting the real and imaginary parts of
the IFRF versus frequency and after curve-fitting them, the modal parameters can
16
be estimated by simple inspection; indeed, the intercept point of the real part with
the real axis provides the natural frequency; either the slope or the ratio of the two
slopes provide both the phase and the magnitude of the modal constant; and any of
the plots provide an estimate of the damping ratio. Dobson also developed a second
method, this time called the “Dobson method” which takes advantage of the complex
modes and corrects for neighboring ones [49]. The aforementioned methods are just to
cite the few most important covering the frequency-domain direct methods. Mendez
Maia in his PhD dissertation [50] presents a very thorough summary of the avail-
able indirect methods which are reproduced here for the reader’s convenience: the
Gaukroger-Skingle-Heron (GSH) [51], Ewins Gleeson [52], Frequency-Domain Prony
(FDPM) [53], the Complex Exponential FD method [54], the Eigensystem Realization
Algorithm in the Frequency Domain (ERA-FD) [55], the Rational Fraction Polyno-
mial (RFP), the global RFP mentioned in [56], the Global Method [57] and the
Polyreference Frequency Domain Method (PRFD) [58].
Among the direct-methods is a smaller number of techniques, both for single-input
single-output (SISO) and multiple-input multiple output (MIMO) systems. These
include the Simultaneous Frequency Domain method (SFD) introduced by Coppolino
[59]; the Identification of Structural Parameters method (ISSPA) by Link [60]; the
Spectral method by Klosterman [61]; and the Multi-Matrix method, developed by
Leuridan [69].
2.2.2 Time-Domain Methods for Modal analysis
So far, a brief description and references of many of the available frequency-domain
modal analysis methods has been provided. On the other hand, several time-domain
methods are also available, some of which have gained great popularity due to cur-
rent computational availability. These methods rely on time response data in the
form of acceleration, velocity or displacement history. It should be noted that within
17
this category, many methods use the information from the inverse fourier transform
(IFFT) to take advantage of the averaging of most of the frequency-domain tech-
niques and thus reducing noise implications. Within this category the most popular
direct methods are the Complex Exponential (CE) for local identification, and the
Least Squares Complex Exponential (LSCE), which is an extension of the CE method
for global identification, both proposed by Spitznogle and Quazi [64]. Vold and Rock-
lin [65] extended the previous method which is for SISO systems to a more versatile
MIMO version called the Polyreference Complex Exponential Method (PRCE), whose
principal improvement was the fact that it could reveal hidden modes due to input
locations near nodal points. Perhaps one of the preferred time-domain methods up-
to-date is the Ibrahim Time Domain Method [66], not only for its integration with the
state-space representation of the equations of motion and consequently with Control
Theory, but also because it was a foundation for the development of further matrix
methods for modal identification such as the ERA. Among the direct methods one can
find a couple of popular ones; namely the Auto Regressive Moving-Average method
(ARMA), based on Gersch’s [67] work; and the Direct System Parameter Identifi-
cation method (DSPI) developed by Leuridan [69]; both of which include statistical
analysis as an assurance criteria.
2.3 Modal Correlation
Modal testing has provided researchers the means of obtaining modal parameters
from experimental information collected from a wide range of setups. However, all
the collected information needs to be compared with the chosen mathematical model
to define, with a fair amount of certainty, the desired behavior of the system. This
comparison between observed results and expected ones is called correlation; in this
particular case, modal correlation is the link that connects analytical and experi-
mental structural dynamics and mechanical vibrations. The most popular modal
parameter correlation is called modal assurance criteria (MAC), which is basically a
18
squared linear regression coefficient based on the Cauchy-Schwartz inequality and was
first proposed by Allemang in 1980 [71]. As Allemang wrote: “The historical develop-
ment of the modal assurance criteria originated from the need for a quality assurance
indicator for experimental modal vectors that are estimated from measured response
functions” [71], and whose intended purpose was orthogonality checks. Since then,
many modal assurance methodologies have been developed depending on the spe-
cific correlation analysis that needed to be performed. These include the coordinate
modal assurance criterion (COMAC), frequency response assurance criterion (FRAC),
frequency-scaled modal assurance criterion (FMAC), partial modal assurance crite-
rion (PMAC), scaled modal assurance criterion (SMAC) and reciprocal modal vector
assurance criterion (RVMAC.) Some of these methods will be briefly described herein.
Validation of experimentally-obtained modal parameters such as mode shape vectors
or natural frequencies can be performed both in vector or scalar fashion; the former,
through calculating the complex modal scale factor; and the latter, through a scalar
measure of vector consistency. The vectorial correlation technique is called modal
scale factor (MSF), whereas the scalar is called modal assurance criterion. These two
correlation approaches are defined in Eqns. (2.5) and (2.6), for two arbitrary vectors
ψa and ψb.
MSF =ψaψbψbψb
. (2.5)
MAC =|ψaψb|2
(ψaψTa )(ψbψTb ). (2.6)
The COMAC utilizes both the mode shape vector and the natural frequency, called
together mode pairs, to verify linear relationship between modes either experimental
or analytical (similar to Chapter 6). However, its direction is more towards identify-
ing which DOF of the measurements contribute negatively to a low value of MAC. It
19
was developed by Lieven [72] in 1988.
Various authors extended this methodology to the frequency-domain. Heylen &
Lammens proposed a similar technique to compare, not modal vectors but frequency
response function vectors using also a squared linear regression coefficient [75]; such
method took the name of frequency response assurance criterion (FRAC), and later
Fotsch & Ewins introduced a frequency scaling upgrade to the MAC method “such
that the mode shape correlation, the degree of spatial aliasing and the frequency
comparison can be displayed in a single plot” [74]. In their proposed methodology,
graphical representations of the mode shape correlations with circles of diameters
equal to the MAC values are compared in a chart such that related modes would
appear in the form of a 45◦ line with big dots and non-correlated modes will appear
as smaller dots spread throughout the plot. Specialized extensions of the MAC are
used commonly when only a desired part of the information is needed. For such cases,
Heylen & Janter proposed the partial modal assurance criterion (PMAC), and the
spatial modal assurance criterion (SMAC) [75]. The former is used to correlate parts
of modal vectors, and it is specially useful when computational efficiency needs to be
achieved by disregarding non-relevant information; the latter compares vector spaces;
that is to say that the SMAC is the least squares solution of a vector transformation
equation as follows:
ψe = ψaQ (2.7)
where Q is the transformation matrix.
SMSF = (ψTa ψTb )−1(ψTa ψ
Tb ). (2.8)
Equation (2.8) represents the least squares solution of Eq. (2.7), minimizing
‖ψb − ψaQ‖.
20
When rotational degrees of freedom are used, a special scaled MAC was proposed
by Brechlin et al., where the weighting matrix is such that it balances the scaling of
translational and rotational DOF’s in the modal vectors [76].
There is also a great deal of research on damage detection approaches using the
Cauchy-Schwartz inequality. Messina proposed an assurance criterion-based method
for detecting damage locations called the Damage Location Assurance Criteria (DLAC),
and later upgraded his method to quantify such damage calling it the Multiple Dam-
age Location Assurance Criteria (MDLAC) [77] [78]. Koh & Dyke extended this work
by adding information from the sensitivity matrix to the formulation and used ge-
netic algorithms in an effort to obtain a computationally-efficient algorithm for vector
comparisons [79]. While both Messina’s and Koh’s work used natural frequency shift
vectors as comparison criteria, Pandey & Biswas used mode shape information to
determine location and amount of damage [80].
It is clear that the method proposed herein has a close relation to those cited
above but with a different scope than to determine damage in structures. However,
the conception of the proposed method in this thesis is quite similar inasmuch as both
approaches propose an iterative comparison of vectors.
In the present thesis, the use of a model-correlation-based method will be used
only for comparison of experimental and analytical mode shape vectors; therefore, to
construct a mapping matrix between analytical and experimental mode shapes. As
many of the MAC formulations use Hermitian matrices instead of transposed ones,
this proposed formulation will deal only with real mode shape vectors. Should a case
where complex modal vectors arise, the formulation must be corrected accordingly.
21
3. THEORETICAL BACKGROUND
3.1 Introduction
The identification method used for the completion of this thesis is peak-picking on
FFTs. Although there is a vast variety of techniques available, such a technique was
selected for its simplicity and ease of application. Also, because the identification
step is not the central topic of the present work. In this section, the theoretical
background of these two methods is explained in detail.
3.2 The Fourier Transform
The Fourier Transform is a basic mathematical resource in signal processing. Its
essential purpose is to decompose any signal or function into sinusoidal components
in an interval from −∞ to +∞. Since real-life signals are finite, a couple of good
artifices are used to transform finite signals into infinite ones: the first is to transform
the finite signal into periodic by adding itself to the right and left of the original
signal; and the second, is to add zeros to the left and right of the original signal.
Consider an arbitrary function x(t), the Fourier transform (FT) is defined as:
F{x(t)} = X(w) =
+∞∫−∞
x(t)e−iωntdt. (3.1)
Note that the FT is a function of the frequency ω. Subsequently, the Inverse
Fourier Transform (IFT) is defined as:
F−1{X(ω)} = x(t) =1
2π
+∞∫−∞
X(ω)eiωntdw. (3.2)
The previous formulation is theoretical and intended for a continuous signal which
is not the case in real life. For discrete signals that occur through sampling, instead
22
of a continuous time t, a time step tk = k∆t is used. The equivalent FT of such
discrete system is:
Xn =N−1∑k0
xke−iωt (3.3)
where ωt = 2πNnk. The sampling frequency is defined as:
fs =1
∆t. (3.4)
Only half of the signal representation by the FT is useful due to the Nyquist-
Shannon theorem which states that the maximum frequency represented by a discrete
signal is equal to half the sampling frequency with which it was generated, i.e.,
fs ≥ 2fc (3.5)
where fs is the sampling frequency, per unit of time (Hz), and fc is the highest
frequency contained in the signal. The inverse Discrete Fourier Transform (IDFT) is
defined as:
xk =1
N
N−1∑n=0
Xnei2π k
Nn. (3.6)
For this thesis, the MATLAB function fft.m and ifft.m are used which are based
on the Fast Fourer Transform, a method that breaks the DFT into smaller size DFT
couples for a faster calculation.
3.3 Frequency Response Function
Consider the single degree of freedom mass-spring-dashpot system depicted in Fig.
(3.1), whose equation of motion is:
mx+ cx+ kx = f(t). (3.7)
The Laplace transform of such a system is:
ms2X(s) + csX(s) + kX(s) = F (s). (3.8)
23
Figure 3.1. Single DOF system.
Isolating X(s), yields:
X(s) = H(s)F (s) =1
ms2 + cs+ kF (s) (3.9)
where m is the mass, c is the damping coefficient, and k is the spring stiffness. The
expression of H(s) in Eq. (3.9) is the frequency response function (FRF) or transfer
function (TF) of the system. Assuming s = iω, one has:
H(iω) =1
mω2 + ciω + k. (3.10)
From this expression, one could obtain the original function very straightforwardly
by applying the inverse Laplace transform.
x(t) = L−1
{1
mω2 + ciω + kF (s)
}. (3.11)
From the Laplace transform tables is known that for a second order system:
L1
{1
(s+ α)(s+ β)
}=
1
β − α[e−αt − e−βt] (3.12)
then,
α =c− i√
4mk − c22m
β =c+ i√
4mk − c22m
.
(3.13)
24
Assuming an under-damped system and an impulse input [f(t) = δ(t) → F (s) =
1],
x(t) =1
m(βα)[e−αt − e−βt]. (3.14)
Finally, defining k = mω2n, c = 2mωnζ, and ωd = ωn
√1− ζ2, the original function
x(t) is:
x(t) =1
mωde−ζωnt[sin(ωnt)]. (3.15)
The FRF describes the response of the system to arbitrary inputs.
3.4 The Peak Amplitude Method
Once the FRF of a system has been determined, and its frequency peaks are vi-
sually available, a straightforward method of modal identification is used: the peak
amplitude method which takes advantage of the property of frequency domain func-
tions to provide a single spike for each frequency content (in ideal cases). After a peak
has been selected, and with the aid of the half-power method, the modal parameters
can be estimated according with the following procedure:
1. The damping ratios can be estimated by evaluating the steepness of each peak,
the steeper the peak, the lighter the damping ratio.
2. The natural frequency of each peak is determined by inspection of the frequency
axis of the FRF plot.
3. The mode shapes can be estimated by the relative amplitudes of peaks at differ-
ent locations of the structure (each sensor determines one location). Therefore,
the greater the resolution, the better the mode shape estimation.
Figure (3.2) shows the half-power technique for modal parameter estimation [84],
where Q represents the amplitude of the peak at the resonance frequency, points R1
and R2 are the half-power points used to define the bandwidth of the system.
Mode shapes are going to be estimated using the complex feature of the FRF.
As a complex function, the imaginary part of such function gives useful information
25
Figure 3.2. The half-power method.
about the phase and relative magnitude of the peaks in each sensor location of the
structure, providing important insights about the mode of vibration at the frequencies
of interest. Consider the FRF matrix row shown in Fig. 3.3 which corresponds to
a cantilever beam with three accelerometers. The relative amplitude of each sensor
along with its phase (above or below the horizontal axis) determine the mode shape of
the first degree of freedom (corresponding to the first peak occurrence) drawn in blue.
If a similar procedure is done for the second, third, and so forth set of peaks, a set of
mode shape vectors is obtained. It must be noted that appropriate resolution has to
be selected to determine as clear and accurate as possible mode shapes and to avoid
aliasing. An alternative technique to the aforementioned is to display the imaginary
part of all the transfer functions generated by each accelerometer in parallel fashion,
26
known as a waterfall plot. This representation allows the analyst a much clearer view
of the big picture for each mode.
Figure 3.3. Estimation of the first three mode shapes of a cantileverbeam with three sensor locations.
27
4. A CORRELATION METHOD
The proposed correlation technique is explained in this chapter. As stated above,
the related mathematical principle used for this development is the Cauchy-Schwartz
inequality, which states that for all vectors x and y of an inner-product space it is
true that:
|〈x, y〉|2 ≤ 〈x, y〉 · 〈y, y〉. (4.1)
Therefore, the relationship between the left side and the right side of the inequality
is always less than or equal to unity.
|〈x, y〉|2
〈x, y〉 · 〈y, y〉≤ 1 (4.2)
which can be used as a suitable correlation index between vectors x and y.
4.1 Correlation based on the Modal Assurance Criterion
The correlation method proposed herein is the application of such inequality which
is a squared linear regression correlation coefficient used originally for orthogonality
checks. This coefficient is very sensitive to large differences between the comparing
vectors (squared error minimization) and consequently, not very sensitive to small
changes. This indicator provides a means for comparing vectors originated in differ-
ent sources: e.g., a FEM-originated modal vector versus an experimentally-originated
one; or any other source. Therefore, no analytical model is needed for this method-
ology to work.
28
The modal assurance criterion is defined as the relationship between the degree
of linearity between two vectors, in this case two modal vectors: one reference and
one experimentally-obtained. The equation for this relationship is:
MAC =|〈ψaψe〉 |2
〈ψaψTa 〉〈ψeψTe 〉(4.3)
where ψa and ψe are the vectors which relation is to be studied, and the T superscript
refers to ‘transposed’. The modal assurance criterion is a scalar value from zero to
one. Zero represents no correlation whatsoever and a MAC value of unity represents
a consistent correspondence. The modal assurance criterion is an indicator of con-
sistency only, that is to say that it will indicate whether two vectors are similar or
not; but it will not indicate if a given vector is correct or valid. Therefore, the use of
an analytical or finite element generated reference vector is always recommended to
provide a valid analysis.
The comparison procedure has two sides: on one side, the first n mode shapes
along with the first n− 1 stacked mode shapes, producing a group of 2n− 1 vectors,
all obtained in an analytical or approximated fashion; on the other, the 2n−1 vectors
obtained experimentally. This will yield a (2n− 1)× (2n− 1) mapping matrix with
elements between zero and unity.
Consider vectors ψ1, ψ2 and ψ3 to be the first three mode shapes of an arbitrary
structure, using r sensors as the experiment’s resolution:
ψ1 =
a1a2...ar
ψ2 =
b1b2...br
. . . ψn =
n1n2...nr
. (4.4)
29
A pre-correlation array is formed by combining the selected number of mode shapes
along with these same mode shapes successively arranged in a stacked fashion: the
first stacked vector will include the first two mode shape vectors; the second stacked
vector will include the first three mode shape vectors, and so forth.
ψ1 =
a1a2...ar
ψ2 =
b1b2...br
. . . ψn =
n1n2...nr
ψn+1 =
a1a2...arb1b2...br
. . . ψ2n−1 =
a1...arb1...br...n1...nr
.
(4.5)
In Eq. (4.5), ψi represents pre-correlation mode shape vectors. The reference array
must have the same structure so a MAC calculation can be performed between the
two to obtain a correlation matrix of the following structure:
CM =
ψ1r1 ψ1r2 . . . ψ1rn 0 0 . . . 0
ψ2r1 ψ2r2 . . . ψ2rn 0 0 . . . 0...
.... . .
......
... . . ....
ψnr1 ψnr2 . . . ψnrn 0 0 . . . 0
0 0 . . . 0 ψn+1rn+1 0 . . . 0
0 0 . . . 0 0 ψn+2rn+2 . . . 0...
... . . ....
......
. . ....
0 0 . . . 0 0 0 . . . ψ2n−1r2n−1
(4.6)
where ψi and ri are the experimental or case study and the reference pre-correlation
vectors, respectively. The upper left region of the matrix corresponds to correlation
between vectors of the same dimension for both the experimental and the reference
modes, whereas the off-diagonal zero regions correspond to cases where no correlation
30
can be checked because of dimension incompatibility. Finally, the diagonal terms after
the ψnrn term correspond to correlation between stacked mode vectors of identical
dimension. A high correlation result will have values close to 1 in the entire CM
matrix diagonal. A simplification of this step is to correlate only stacked vectors,
leaving out of the analysis the single mode shape vectors as the latter, when analyzed
individually, produce poor results due to individual mode shape similarity. More-
over, when comparing stacked-modes vectors only, the correlation matrix turns into a
correlation vector much more straightforward to analyze from the analyst standpoint.
The results of the correlation matrix or vector can be presented by either the
matrices themselves or bar plots, whether presented in 3D for the case of correlation
matrices and in 2D for the case of vectors, which are much more easy to understand.
An example of this plot for a 5 × 5 successful correlation matrix is presented in
Fig. 4.1.
12
34
5
1
2
3
4
5
0
0.2
0.4
0.6
0.8
1
Experimental modeReference mode
Co
rrel
atio
n
Figure 4.1. An example of a 5 × 5 correlation matrix.
31
Model correlation matrices are very common when a mode needs to be checked for
validity in cases of computational or torsional modes. For this particular project, the
benefits of the modal assurance methodologies are used with a distinct objective: to
implement a methodology for verifying the closeness of a model boundary condition
parameters to those of an experimental result. This is why the method will be further
referred to as “MAC-based” instead of modal assurance criteria.
4.2 Resolution Study
As stated previously, the resolution chosen for data acquisition is of crucial im-
portance in the stability of the method. Indeed, the method will become less reliable
with the reduction of number of sensors; on the contrary, the higher the resolution,
the better correlation results one can obtain. Three runs of the algorithm were made,
using sixteen, eight and four accelerometers to determine a pattern of correlation.
These analyses had to be done with analytical data to determine the sensitivity of
the method to the number of sensors used without the contamination of data with
external factors such as noise. The results are presented Figure 4.2, the correlation
value for one of the cases where both vectors are analytical is plotted against number
of accelerometers. It can be clearly observed how the correlation drops dramatically
when less than eight accelerometers are used for the case of a beam of 4572 mm
length. This is also confirmed by a poor approximation in the finite element model.
Setting a general rule of thumb, the stability limit occurs when the length of
discretized elements of the beam is at least 1/10 of the length of the beam, so
LbeamLelements
≤ 10. (4.7)
32
Figure 4.2. Resolution sensitivity of the proposed method.
PART II: NUMERICAL VERIFICATION
33
5. MATHEMATICAL MODELS
Detailed derivations of the mathematical models that were used for this work will be
described herein. Both the analytical and, though approximated, the finite element
method will provide insights of the “exact” solutions. Both methods will be used
to prove that either approach could be used as a correlation comparison reference.
This is useful specially when an analytical model cannot be obtained and a finite
element approximation has to be used. Chapter 6 in addition will address the use
of the method with data gathered from the models developed in this chapter. The
methodology used will be thoroughly explained.
5.1 Analytical Models
The experimental setup which will be explained in detail in Chapter 7 consists of
a beam-like bridge structure with four parallel beams connected by common flanges
at set distances from the boundaries for simultaneous vibration of the structure. This
structure was intended to behave like a continuum beam subjected to different types
of boundary conditions such as pivoted (pinned), fixed (clamped), elastic (spring), or
any combination of these. It should be noted that free, fixed and pivoted boundaries
are commonly referred to as “classical” boundary conditions, whereas any other form
is known as “non-classical”. Since the purpose of this thesis is not to present extensive
mathematical derivations of the solutions of all the aforementioned cases, three cases
have been selected to be developed just to demonstrate the validity of the method:
these cases are the more general pinned-pinned or simply-supported beam, a beam
with torsional springs on both ends, and a beam with rotational masses on both ends.
34
Consider the two most classical cases of boundary conditions for a beam, depicted
in Fig. 5.1(a) for a simply-supported beam case (pinned-pinned), and Fig. 5.1(b) for
a clamped-clamped case.
(a) Simply-supported.
(b) Clamped-clamped.
Figure 5.1. Classic boundary conditions.
These two cases can be viewed as special cases of a beam supported both on
rotational and linear springs at ends. Classical boundary conditions are those of
more extensive use in engineering and the bibliography; namely, pinned, free and
clamped. However, these are just particular cases of beams on elastic supports where
specific values of the support spring stiffnesses are used. For instance, a beam with
pinned ends and torsional supports of zero stiffness is actually a simply-supported
beam.
35
5.1.1 Simply-Supported Beam Model
Consider a continuous beam of homogeneous material; modulus of elasticity E;
second moment of area Ix; density ρ; length L; and uniform cross-sectional area A.
The governing differential equation of such a beam according to the Euler-Bernoulli
theory of beams can be written as:
−∂2w(x, t)
∂t2=EI
ρA
∂4w(x, t)
∂x4. (5.1)
For simplicity, the effects of shear, stress and rotational inertia have been neglected
in this particular case. Since Eq. (5.1) is a partial differential equation, it can be
solved by the technique of separation of variables; hence, this equation has a solution
of the form,
w(x, t) = q(t)φ(x) (5.2)
for a single mode, where q(t) is the frequency equation and φ(x) is the mode shape
equation. Substituting Eq. (5.2) into Eq. (5.1) yieds,
−φ(x)d2q(t)
dt2=EI
ρAq(t)
d4φ(x)
∂x4(5.3)
and rearranging,
− 1
q(t)
d2q(t)
dt2=EI
ρA
1
φ(x)
d4φ(x)
∂x4. (5.4)
Solving this equation requires both sides to be equal to a constant which must be pos-
itive. This constant is denoted ω2 and is the frequency parameter of the system. Now
the expression can be separated into two ordinary homogeneous differential equations
each one of which depends only on one variable as follows,
d2q(t)
dt2+ ω2q(t) = 0 (5.5)
andd4φ(x)
dx4− ω4φ(x) = 0 (5.6)
36
with a frequency parameter defined as
β4 =ρAω2
EI. (5.7)
The general solution for Eq. (5.5) is of the form,
q(t) = cos(ωt− α). (5.8)
Similarly, the general solution for Eq. (5.5) is
φ(x) = Asinβx+Bcosβx+ Csinhβx+Dcoshβx. (5.9)
The interest in the present analysis lies in Eq. (5.7) and Eq. (5.8). The solution of
the latter, though straightforward, requires a previous determination of parameter β
from the former. To complete the problem statement, boundary conditions must be
defined to solve for the specific case of interest. This is a cumbersome task that re-
quires careful mathematical manipulation along with a correct definition of boundary
conditions. The boundary conditions for a simply-supported beam can be expressed
as
φ(x = 0) = φ(x = 0) = 0
φ′′(x = 0) = φ′′(x = L) = 0.(5.10)
Applying boundary conditions in Eq. (5.10) for the case of x = 0 into Eq. (5.9)
produces,
37
φ(x = 0) = Asin 0 +Bcos 0 + Csinh 0 +Dcosh 0 = 0 (5.11)
and
φ′′(x = 0) = β2(−Asin 0−Bcos 0 + Csinh 0 +Dcosh 0) = 0 (5.12)
thus giving,
B +D = 0
−B +D = 0
⇒ B = D = 0.
(5.13)
Similarly, applying the boundary conditions corresponding to x = L produces an
homogeneous system of equations,
AsinβL+ CsinhβL = 0
−AsinβL+ CsinhβL = 0.(5.14)
For a non-trivial solution to exist, the determinant of the coefficients must vanish,
∣∣∣∣∣∣ sinβL sinhβL
−sinβL sinhβL
∣∣∣∣∣∣ = 0 (5.15)
which produces the following requirement,
sinβLsinhβL = 0. (5.16)
The only possibility for Eq. (5.16) to be zero is when the values of βL are equal
to nπ (n = 1, 2, 3, ...). Now, the equation of mode shape along with the frequency
parameter have been completely derived and can be expressed as,
38
φ(x) = sinβx (5.17)
and
β =nπ
L. (5.18)
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1−1
−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
1
Am
plitu
de
x/L
1st2ND3RD4TH
Figure 5.2. First four modes of a simply-supported beam.
This derivation produces the well known mode shape for a simply-supported beam
shown in Fig. 5.2. As expected, the shape is that of a sine function whose frequency
increases by a factor of one in each resonant frequency. It should be noted that at the
boundaries, no slope variation is observed as no constraint is acting on those ends.
5.1.2 Beam with a Torsional Spring Support at Each End
Consider now the case of a beam shown in Fig. 5.3. The material properties are
exactly the same as in Sec. 5.1.1, with the addition of two torsional springs, whose
stiffnesses are kθL and kθR for the left and right side, respectively
The boundary conditions for this case are,
φ(x = 0) = φ(x = L) = 0 (5.19)
39
Figure 5.3. Beam on elastic supports.
for the displacement at both ends, and,
d2φ(x = 0)
dx2=KθL
EI
dφ(x = 0)
dx
d2φ(x = L)
dx2= −KθR
EI
dφ(x = L)
dx
(5.20)
for the bending moment at both ends. Substituting the boundary condition equations
into the general solution [Eq. (5.9)] and after some manipulation, an expression of
the frequency equation can be obtained:
K∗21 +K∗1β(1 + α)(sinβcoshβ − cosβsinhβ)
α(1− cosβcoshβ)+
2β2sinβsinhβ
α(1− cosβcoshβ)= 0 (5.21)
where K∗1 = KθL × 180π
, K∗2 = KθR × 180π
and α =K∗
2
K∗1, (0 < α < 1).
The mode shape equation is determined,
φ(ξ) = sinβξ − sinhβξ + γ
[cosβξ − coshβξ − 2β
K∗1sinhβξ
](5.22)
where
γ =sinhβ − sinβ
cosβ − coshβ − 2βK∗
1sinhβ
. (5.23)
40
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1−1
−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
1
Am
plitu
de
x/L
SS−1stCC−1STSS−2NDCC−2NDSS−3RDCC−3RDSS−4THCC−4TH
Figure 5.4. Limiting mode shapes for vibrations of a beam on elas-tic supports (solid lines: simply-supported; dashed lines: clamped-clamped).
This constitutes an intermediate case between the simply-supported and the clamped-
clamped cases. Fig. 5.4 shows these limiting cases, the solid line represents the simply-
supported case where both spring constants Kθ1 and Kθ2 are equal to zero; similarly,
the dotted line represents the clamped-on-both-ends case where the spring constants
are equal to infinity. Any combination of spring constants at left and right ends will
yield modal parameters within these two limits, being the upper limit the clamped-
clamped case; and the lower limit, the pinned-pinned case. This principle also applies
for the natural frequencies, as the highest will occur for the clamped-clamped model;
and the lowest, for the simply-supported model. All natural frequencies for spring
supported beams will be between these two bounds.
5.1.3 Beam with Rotational Masses on Both Ends
The third model considered in the present thesis, mainly due to discrepancies in
the observed modal parameters versus the ones predicted by the previous two models
41
is a simply-supported beam with the addition of rotational inertial masses on both
ends. Consider a beam similar to that shown in Fig. 5.5
Figure 5.5. Simply-supported beam with rotational masses on both ends.
For this case, the first two boundary conditions are the same as a simply-supported
case [Eq. (5.19)], whereas the second two include the effect of the rotary inertial
masses:
d2φ(ξ = 0)
dξ2=−Io1ρAL3
β4dφ(ξ = 0)
dξ
d2φ(ξ = 1)
dξ2=
Io2ρAL3
β4dφ(ξ = 1)
dξ
(5.24)
where ξ is the non dimensional distance x/L. Again, solving the system of equa-
tions produced by these boundary conditions will end up giving the transcendental
eigenvalue equation along with the mode shape equation:
I∗21 + I∗1β3(1 + α)(sinβsinhβ − sinβcoshβ)
2αsinβsinhβ+β6(1− cosβcoshβ)
2αsinβsinhβ= 0 (5.25)
where I∗1 = Io1 × 180π
, I∗2 = Io2 × 180π
and α =I∗2I∗1
, (0 < α < 1).
The mode shape equation is:
42
φ(ξ) = sinβξ − sinhβ
sinhβsinhβξ + γ
[cosβξ − coshβξ +
(coshβ − cosβ)
sinhβsinhβξ
](5.26)
where
γ =(sinβ/sinhβ)− 1
[(coshβ − cosβ)/sinhβ]− (2I∗1/β3). (5.27)
Such equation has different limiting cases than those of the spring supported case
(Eq. 5.4). Calculation of such limiting cases has to be performed by combining the
knowledge of the current mode shape with that of previous ones; i.e., when considering
the value of n for the calculation of the mode shape, some of the frequencies to consider
are of previously obtained cases (n− 2). This is a rather delicate procedure as errors
can be introduced very easily and will be explained in detail in the finite element
section. The first four mode shapes for this case are shown in Fig. 5.6.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1−1
−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
1
x/L
Am
plitu
de
1st2nd3rd4th
Figure 5.6. First four mode shapes of beam with rotational masses on both ends.
Note the similarity between the first and the third mode shapes; and between the
second and the fourth mode shapes.
43
5.2 Finite Element Models
A second class of mathematical model developed for this project is a finite element
approximation. A simple finite element model (FEM) was created using a toolbox
developed at the Intelligent Infrastructure Systems Lab by Caicedo [4] using MAT-
LAB [81]. Such a model was constructed with as few nodal points as possible to
obtain acceptable results within a reasonable tolerance without significant computa-
tional demand. The purpose of building this model was simply to have two sources
of analytical reference given that a FEM can be treated as analytical.
It should be clearly noted that the development of this type of model is a mere
idealization of the real structure that will be introduced and explained in Chapter 7
and throughout this chapter, it will be considered as such.
To simulate the responses of the selected models to a variety of boundary condi-
tions and material properties, the finite element model created consists of 72 Euler-
Bernoulli beam elements connecting 76 nodes. Each beam of the structure is dis-
cretized in 17 Euler-Bernoulli beams with 18 nodes along its length. Such a model
has a total of 53 degrees of freedom, provided that each node is defined with six
degrees of freedom: three for translation and three for rotation, some of which are re-
stricted to ensure a 2D behavior of the vibrations. Further discussion of the creation
of the finite element model are include in the following sections.
5.2.1 Simply-Supported Beam Model
The finite element model of the structure for the simply-supported case was de-
veloped using 76 nodes and 72 beam elements. Each beam was discretized in 18
elements to replicate as close as possible the experimental setup used, which had 16
accelerometers in each beam. Since the pursued goal in this project is to study the
modal behavior of the structure, the finite element toolbox used calculates simple
44
mass and stiffness matrices assembled from the conditions given in the input data
which include:
1. Material properties such as density, modulus of elasticity and shear modulus.
2. Section and geometric properties like cross-sectional area, second moment of
area in x and y directions and the St. Venant’s rotational constant.
3. Nodal coordinates defined by the number of nodes chosen and their coordinates
in space expressed as a 4-element vector ([node # x coord y coord z coord]).
Since this is a basic toolbox, discretization has to be defined manually.
4. Boundary conditions. The toolbox calculates all the parameters using 6 degrees
of freedom per node corresponding to three translational DOFs and three rota-
tional DOFs for a total of 3 directions (x, y and z). It was of great aid to select
the desired DOFs accordingly to avoid movement in unwanted directions.
5. Structural information like rigid links, lumped masses and springs.
After defining all the properties listed above, the process of construction of the
model is straightforward, due to the nature of the toolbox being object-oriented.
This means that what needs to be defined are objects like the material, sections,
connectivities, links, masses, but all these defined as independent objects stored as
such, instead of programmatic variables. A short example of definition of material
parameters is given below:
E = 200e9; % Modulus of Elasticity [Pa]
nu = 0.3; % Poisson coefficient
L = 4.572; % Beam length [m]
Iy = 5.036e-5; % 2nd moment of area [m^4]
Ix = 1.550e-6; % 2nd moment of area [m^4]
rho = 7850; % Material density [kg/m3]
A = 0.004; % Cross section area [m^2]
45
G = E/(2*(1+poisson)); % Shear modulus [Pa]
J = 2.57e-5; % Torsional constant [m^4]
Section and material definition examples are also provided:
%%% Define Section(s) %%%
section = section(’bridge’,A,Ix,Iy,J);
In this definition, a section called ’bridge’ is created and its properties are a section
area A, second moments of inertia Ix and Iy, and torsional constant J .
%%% Define Material(s) %%%
steel = material(’steel’,E,G,rho);
This definition on the other hand, creates a material called ’steel’, whose properties
are a modulus of elasticity E, a shear modulus G, and a material density of ρ.
Sections and materials can be created as needed for a given structure and assigned
to different elements of the model, making it quite versatile at the moment of modeling
complex structures. With all the objects defined, the toolbox calculates the mass and
stiffness matrices and therefore, through solving the eigenvalue problem, the modal
parameters are obtained. Fig. 5.7 shows the nodal definitions of the structure.
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
0
0.5
−0.2
0
0.2
1 2
3 4
5 6
7 8
9 10
11 12
13 14
15 16
17 18
19
20 21
22 23
24 25
26 27
28 29
30 31
32 33
34 35
36 37
38
39 40
41 42
43 44
45 46
47 48
49 50
51 52
53 54
55 56
57
58 59
60 61
62 63
64 65
66 67
68 69
70 71
72 73
74 75
76
y
x
Figure 5.7. FEM node layout.
46
Figure 5.7 shows only the nodal points on the beams of the structure, as only these
are of interest for the present analysis. Should a future need to study the implications
of the piers or other connections present, these could be included without problems.
The model is defined completely when the boundary conditions are set. Indeed, this
is a boundary value problem. For the simply-supported case, the boundary conditions
are set as,
y(nodes 1, 20, 39 58) =
tx = constrained
ty = constrained
tz = constrained
rx = constrained
ry = free
rz = constrained
=
1
1
1
1
0
1
(5.28)
where tx, ty and tz are the translational DOF’s along the subscripted axis, and rx, ry
and rz are the rotational DOF’s around the subscripted axis. The result of this def-
inition is such that nodes 1, 20, 39 and 58 will be restricted to translate in the three
directions, and to rotate around x and z directions; whereas they will be free to rotate
around the y axis. This is consistent with a pinned boundary.
Data entering for modeling in the fem toolbox uses a similar fashion as that shown
in Eq. 5.28 in the sense that every element that contributes any effect on a node will
have a 6-element vector associated with it. For instance, a nodal mass will have three
translational and three rotational directions; a spring element acting over a node will
have three translational and three rotational stiffnesses, and so on.
The results of the simulation using finite elements is shown in Table 5.1 for the
first five natural frequencies. As expected, there is very good concordance between
the approximate (FEM) frequencies with those of the analytical model.
For the case of the mode shapes, the output of the program is a 3D plot of each
mode. For simplicity and clarity, only three 2D mode shapes are shown in Fig. 5.8.
47
Table 5.1. FEM frequency comparison (in Hertz).
Analytical FEM
7.4869 7.4869
29.9476 29.9479
67.3821 67.3856
119.7905 119.8100
187.1726 187.2464
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1−1
−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
1
x/L
Am
plitu
de
Figure 5.8. Simply-supported case mode shapes generated by FEM.
5.2.2 Beam With a Torsional Spring Support at Each End
In a similar way, the finite element method was used for modeling a beam with
torsional springs at both ends. This study was done for comparing the behavior of
a beam constrained in an unknown fashion at the borders. Such constraints were
conceptualized as torsional springs that would prevent the system to vibrate freely.
This model results in a very useful general tool to understand the vibration of similar
systems from the simply-supported case, to the clamped-clamped case. Indeed, the
former is nothing but a special case of a beam with torsional springs at both ends
48
whose stiffnesses are zero; whereas the latter is the case of a beam with springs’
stiffnesses close to infinity. Any other combination of stiffnesses will lay between
these two. Again, the natural frequencies and mode shapes for an intermediate value
of stiffness (between zero and infinity) will be presented shortly.
The structure was discretized in a similar way as in the previous section, with the
only difference being that additional nodes had to be declared for defining the springs.
Springs are defined as extra elements connecting fictitious orthogonal nodes to those
already existing. For the case of this investigation, springs were created for the
boundaries only. Figure 5.9 depicts the element configuration of this model, including
the spring elements acting over the boundaries which correspond to the vertical green
lines numbered 1-8, whereas the horizontal lines between dots represent the beam
elements. Rotational springs are represented this way in the model but this does not
mean that they are physically as shown. This type of element is in fact conceptualized
as a box having shear resistance across the faces other than the axial face. An example
of the definition of these elements is:
k(nodes 1, 20, 39 58) =
txtytzrxryrz
=
0000ki0
(5.29)
where ki is the stiffness of the spring in N m/rad. The stiffness used for this example
is k1 = 75000 Nm2 which had to be transformed to units of stiffness per degree of
rotation by multiplying it by 180 /π and k2 = 45000 Nm2 which is 60 % of the oppo-
site stiffness. This difference of stiffnesses is evident when looking at the mode shape
plot and by noticing that the maximum and minimum points are slightly shifted to
the right due to a higher constraint at the left side causing an unsimmetric shape.
If the springs were considered of equal stiffness, the plot would have been perfectly
symmetric.
49
The boundary conditions for this model are similar to those of the simply-supported
with the difference that they have an additional effect imposed by the springs.
00.5
11.5
22.5
33.5
44.5
0
0.5
−1.5
−1
−0.5
02
18
4
x
36
6
17
54
8
35
72
16
53
34
71
15
52
33
70
14
51
32
69
13
50
31
68
12
49
30
67
11
48
29
66
10
47
28
65
9
46
27
64
8
45
26
63
7
44
25
62
6
43
24
61
5
42
23
60
4
41
22
59
3
40
21
58
2
39
20
57
1
38
1
19
56
3
37
y
5
55
7
Figure 5.9. Layout of elements generated by the FEM.
The natural frequencies obtained versus those of the mathematical model are
shown in Table 5.2, and the first three mode shapes in Fig. 5.10. Note that the
mode shapes are somewhat similar to those of a clamped-clamped case which is ex-
pected, since the boundaries indeed have constraints in rotation.
Table 5.2. FEM frequency comparison for a system with rotationalsprings on both ends (in Hertz).
Analytical FEM
15.4534 15.7484
42.9335 43.6126
84.7279 85.8527
140.8820 142.4655
167.3537 213.5939
50
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1−1
−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
1
x/L
Am
plitu
de
Figure 5.10. Elastic support case mode shapes generated by the FEM.
5.2.3 Beam With Rotational Masses on Both Ends
The third finite element model constructed is the counterpart of the theoretical
simply-supported beam with rotational masses on both ends. Given the nature of the
experimental setup which will be explained in detail in Chapter 7, a structure was
built in such a way that two torque arms were attached to the boundaries so that a
boundary-varying structure could be achieved. These arms can be fixed to produce a
clamped boundary; or can be let free to produce a pinned boundary. Unfortunately,
they add a considerable amount of rotary inertial mass to the boundary, changing the
properties of the intended simply-supported configuration. This is the reason why a
third model had to be studied. Consider the beam shown in Fig. 5.5. By inspection of
the experimental structure shown in Fig. 7.3 the most noticeable similarity between
these two pictures is the big attachments at both ends. For the experimental case,
it was not very difficult to determine the inertia effect of these elements, as they are
produced by plates rotating along their edges. The rotational moment of inertia for
this case is,
I0 =1
3mL2 (5.30)
51
where m is mass of the plate, L is arm length and I0 is the rotational inertia. The
masses of the attachments were estimated through the CAD software used for de-
signing the structure and are m1 = 28 kg, and m2 = 14 kg which will produce
two rotational moments of inertia, one for each end, whose units are kg m2 which
were then transformed to units of inertia per units of rotation by multiplying them
by 180/π, and then transformed to the dimensionless rotational moment of inertia
constant,
I∗ =ρAL3
I0. (5.31)
For the setup of study, these values are I∗1 = 54, and I∗2 = 27. The finite element
model was constructed using the same number of nodes and elements as the two
previously described. An internal condition stating the rotational mass effects was
included for each end which will not vary the node and element layouts shown in
previous figures. Boundary conditions were set similarly as these cases. The results
are shown in Table 5.3.
Table 5.3. FEM frequency comparison for a system with rotationalmasses on both ends (in Hertz).
Analytical FEM
5.7413 5.7851
13.6967 13.9492
25.0677 25.2683
50.1448 50.3739
93.4043 93.5265
It is evident how close these frequencies are, confirming that this finite element
model is correctly constructed. Some differences between the values could be caused
52
by the difficulty that posed to estimate accurately the type of rotational moment
of inertia due to the complexity of the structure. For simplicity this parameter was
considered as a rotating plate along its edge, however, some other shapes and off-axis
considerations were not taken into account. Fig. 5.11 shows the first three modes of
the structure.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1−1
−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
1
x/L
Am
plitu
de
1st2nd3rd4th
Figure 5.11. Modeshapes of a simply-supported beam with rotationalmasses on both ends, generated by the FEM.
Three analytical models and three finite element models have been developed.
These were chosen as the source of the model space which will be compared with the
“unknown” experimental model to determine the closest correlation.
5.2.4 Finite Element Model Details
The finite element formulation used in the present thesis calculates the stiffness
and mass matrices from nodal information in which each node has six degrees of
freedom. Figure 5.12 shows a detail of the six degrees of freedom for each node. In
this figure, plain numbers represent nodes, boxed numbers represent elements, R1 to
R6 represent the six DOF for the first node; and R7 to R12, the six DOF for the
second node. Therefore, the global mass and stiffness matrices are of size 6n × 6n
53
where n is the number of nodes, and r are the the DOFs constrained by the boundary
conditions.
Figure 5.12. Beam element with element coordinates.
Mass and Stiffness Matrices
It is assumed that vibration only occurs in 2D; therefore, the unwanted DOF are
constrained in such a way that only the desired directions of movement are char-
acterized by the finite element analysis, reducing the original problem to a 3 DOF
per node: one for translation in the vertical z direction, one for translation in the
horizontal x direction, and one for rotation around the y direction.
Consider a beam element. The stiffness influence coefficient kij is the force in DOF
i due to a unit displacement or unit rotation in DOF j. From the principle of virtual
displacement, the general equation for kij is:
54
kij =
L∫0
EI(x)ψ′′i (x)ψ′′j (x)dx (5.32)
which can be evaluated analytically for i, j = 1...6, resulting in element stiffness
matrix:
ke =EI
L3
L2AI
0 0 −L2AI
0 0
0 12 6L 0 −12 6L
0 6L 4L2 0 −6L 2L2
−L2AI
0 0 L2AI
0 0
0 −12 −6L 0 12 −6L
0 6L 2L2 0 −6L 4L2
(5.33)
where E is the modulus of elasticity, A is the cross-sectional area, I is the area mo-
ment of inertia and L is the length.
Similarly, the mass influence coefficient mij is the force in DOF i due to unit linear
acceleration or unit angular acceleration in DOF j. Applying again the principle of
virtual displacement, the general equation for mij is:
mij =
L∫0
m(x)ψi(x)ψj(x)dx (5.34)
which can be evaluated analytically for i, j = 1...6, resulting in element mass matrix:
me =mL
420
140 0 0 70 0 0
0 156 22L 0 56 −13L
0 22L 4L2 0 13L −3L2
70 0 0 140 0 0
0 54 13L 0 156 −22L
0 −13L −3L2 0 −22L 4L2
(5.35)
where m is the total mass of the element.
55
The 6 × 6 element stiffness matrix ke and the 6 × 6 element mass matrix me
are transformed in global element coordinates through an appropriate transformation
matrix, A [85].
K = Aki (5.36)
M = Ami. (5.37)
Finally, the natural frequencies and mode shape vectors are obtained through
solving the eigenvalue problem for the eigenvalues and eigenvectors, respectively.
Kφ = ω2Mφ (5.38)
Interpolation functions
The displacement of the beam element shown in Fig. 5.12 is related to its six DOF
through:
6∑i=1
ui(t)ψi(x) (5.39)
where ψ is the displacement of the element due to unit displacement. The 6 × 6
stiffness and mass matrices of Eqs. 5.33 and 5.35, respectively turn into 4 × 4 local
element matrices when eliminating the coefficients associated with axial DOF at each
node. Therefore, ψi(x) satisfies the homogeneous boundary conditions:
i = 1 : ψ1(x = 0) = 1, ψ′1(x = 0) = ψ1(x = L) = ψ′1(x = L) = 0 (5.40a)
i = 2 : ψ2(x = 0) = 1, ψ′2(x = 0) = ψ2(x = L) = ψ′2(x = L) = 0 (5.40b)
i = 3 : ψ3(x = L) = 1, ψ′3(x = 0) = ψ3(x = 0) = ψ′3(x = L) = 0 (5.40c)
i = 4 : ψ4(x = L) = 1, ψ′4(x = 0) = ψ4(x = 0) = ψ′4(x = L) = 0 (5.40d)
56
Assuming no shear effects, the governing equation for a beam loaded at its ends is:
EId4u
dx4= 0 (5.41)
whose general solution is of the form,
u(x) = a1 + a2
(xL
)+ a3
(xL
)2+ a4
(xL
)3. (5.42)
After enforcing the boundary conditions of Eqns. 5.40a, constants a1, a2, a3 and
a4 are determined.
ψ1(x) = 1− 3(xL
)2+ 2
(xL
)3(5.43)
ψ2(x) = L(xL
)− 2L
(xL
)2+ L
(xL
)3(5.44)
ψ3(x) = 3(xL
)2− 2
(xL
)3(5.45)
ψ4(x) = −L(xL
)2+ L
(xL
)3(5.46)
which can be used as interpolation functions for the finite element model.
Additional Elements
Once the nodal contributions to the global mass and stiffness matrices have been
established, the interaction between such nodes was defined through rigid links. Four
rigid links were defined at equidistant lengths along the beam; namely, at x = 0,
x = L/3, x = 2L/3 and x = L. The definition pattern for these rigid links were to fix
the node on the front element as master node, and link the corresponding nodes on
the other beams as slave nodes. The rigid link diagram is shown in Fig. 5.13, where
M refers to master node; and S, to slave node.
57
Figure 5.13. Detail of the FEM rigid links.
Notice that these rigid links are not defined as beam elements, therefore they don’t
have any contribution to the overall mass and stiffness matrices.
The boundary condition definition was briefly explained in Sec. 5.2.1. The finite
element toolbox used in the present thesis has the capability to create a boundary
condition in the same way as a nodal DOF is defined, that is to say, a vector of
six elements (three for translation and three for rotation) is defined with values of
zero and one (zero for free and one for constrained). Figure 5.14 depicts an example
definition of boundary conditions for a pinned-free beam element. The left side has
a pinned boundary with free rotation only around the y axis and all other DOFs
constrained; and the right side, has a free translation along the x and z axis, free
rotation around the y axis and all the other DOFs are constrained.
Similarly, lumped masses and spring elements are defined by 6-element vector. For
the former case, three translational mass-contributions and three rotational inertia-
contributions have to be defined; and for the latter, three linear spring stiffnesses and
three rotational spring stiffnesses have to be defined.
58
Figure 5.14. Boundary condition definition example.
The rotational mass FEM explained in Sec. 5.2.3, had only one rotational compo-
nent acting around the y axis, which is consistent with a 2D model. Lumped masses
on each nodal point were neglected. The mass definition is shown in Fig. 5.15. On the
other hand, the spring supported model had rotational springs acting in the boundary
nodes around the y axis, as element that prevented the nodes to rotate freely. As
explained in Sec. 5.2.2, springs are conceptualized as boxes having shear resistance
across the faces other than the axial face.
Figure 5.15. Rotational mass definition example.
59
6. NUMERICAL VERIFICATION
6.1 Introduction
As stated previously, the development of the mathematical models used in this
thesis is the first part of a much broader procedure which is the numerical verifica-
tion of the proposed technique.
A model correlation method has been proposed but before it can be implemented
with experimental data, it has to be tested with unpolluted data so that it can prove
its reliability. In the present chapter, this verification will be performed in an iterative
way so that it can be further implemented with real data. Two sources of data can
be used instinctively from the mathematical models already discussed: the analytical
and the finite element models. It is expected that for the analytical case, a perfect
correlation occurs; whereas for the finite element case, which is an approximation, a
close-to-perfect-correlation occurs.
6.2 Methodology
A good approximation of the structural model was made in Sec. 5.2.3. This case
will be used to run the correlation procedure against the three developed mathemat-
ical models, each one with different parameter values. This constitutes the ‘model
space’, a group of distinct parametrized models to which the experimental case is
going to be compared until the best match is determined. The mathematical models
to be used are:
1. Simply supported beam with both ends free for rotation.
2. Simply supported with:
60
• Rotational springs on both ends with three different kθi values on one end,
and five proportional values going from 20% to 100% of kθi on the other
end. This gives a total of twenty five cases.
• Rotational inertial mass on both ends, with five different Iri values on one
end, and five proportional values going from 20% to 100 % of Iri on the
other end. This gives a total of twenty five cases.
3. Clamped boundary on both ends (clamped-clamped beam.)
The described list contains roughly 80 different analytic models to compare from.
It is clear that the solution space can be as wide as needed to determine the closest
parameter correlation. The procedure to simulate the unknown boundary condition
case is deatiled below:
1. Define the number of modes n to be simulated.
2. Obtain the eigenvalues, ωn and mode shape vectors φn from the analytical
equations or FEM.
3. Aproximate each mode to a state-space model corresponding to the associated
degree of freedom.
4. Obtain the response of the system to a simulated input signal for each degree
of freedom by solving the equations of motion. This constitutes the simulated
data.
5. Apply an identification technique (FFT or ERA) to generated data to obtain the
modal parameters. This step is to replicate a similar procedure of experimental
modes. Noise or any other disturbance can be added in this step.
6. Compare the obtained mode shapes to those of the analytical models using the
correlation method.
7. Plot the correlation matrix in a bar plot to determine the highest value and,
therefore, the best model match.
61
6.3 Model Selection
Figure 6.1 presents five subplots, three for individual and two for combined (stacked)
mode shapes. In each subplot, four curves are overlaid which correspond to the sim-
ulation, the simply-supported, the clamped-clamped and the rotational mass models.
By inspection of this figure, it is evident that there are some lines not matching in
all five subplots. In the figure, the red dotted lines with squared markers represent
the mode shape of the case study; i.e., the simulation obtained from an analytical
equation which is considered the “experiment”. The continuous blue lines represent
the analytical model of the same case but obtained through the FFT process to recre-
ate the method as close as possible to the real experiment. The dashed black and
dashed green lines correspond to the simply-supported and clamped-clamped cases,
respectively.
In the first subplot, all four lines are almost coincident and a no conclusions can be
made to determine which model is closer to the simulated case. The second subplot
shows an almost collinear behavior between the simulation and the rotational mass
models, whereas a slight deviation of the simply-supported and clamped-clamped
models from that of the simulation. Although a deviation is observed, no conclusions
can be made at this point either. The third subplot shows a significant deviation
of the simply-supported and clamped-clamped models from that of the simulation
and, at the same time, it shows how a consistent convergence between the simulation
and the rotational mass analytical model. The last two subplots show a comparison
between models but using stacked vectors instead of individual vectors; for these two
cases, a consistent collinearity is observed between the rotational mass model and the
simulation. Three modes have been used for explaining this step. Depending on the
necessity, more mode shapes can be used to have a better convergence, depending
on the complexity of the model and the availability of good experimental high mode
shapes.
62
Although the previous step produces a fair amount of information about which
model to select, it does not provide a parametric result regarding the boundary con-
ditions of such a model. The goal of this technique is to determine with a great deal
of certainty the parameters involved in the chosen model. For example, if a spring-
supported model is selected, the method would provide an interval of spring-stiffness
values as parameters of the system.
0
0.5
1
mo
de1
−1
0
1
mo
de2
−1
0
1
mo
de3
−1
0
1
stac
k1
−1
0
1
stac
k2
SS CC ROT−MASS MODEL SIMULATION
Figure 6.1. Modal vector comparison.
6.4 Correlation Analysis
Once the visual inspection has been conducted, the ultimate mathematical veri-
fication is done through the correlation methods described in Chapter 4. Consider
again the simulated experiment just detailed in the previous section. Suppose that
some given mode shape vectors have been obtained after performing impact testing to
a structure. Such vectors are shown in Fig. 6.2. This is going to be the experimental
“case study”. Suppose also that the only assumption made for this mode shape is
that it corresponds to the vibration of a beam of known dimensions and material con-
stants but unknown boundary conditions (which is usually the case); i.e., parameters
63
M, C and K have been experimentally obtained. The proposed technique’s aim is
to compare this case study vector with several analytical model cases to determine
which resembles the closest to it.
0
0.5
1
mo
de1
−1
0
1
mo
de2
−1
0
1
mo
de3
−1
0
1
stac
k1
−1
0
1
stac
k2
SIMULATION
Figure 6.2. Case study.
After comparing this mode shape with the classical boundary-condition models
similar to the step described in Sec. 6.3 one can conclude preliminarily that the
obtained mode shape corresponds to a beam with rotational masses on both ends.
However, this is still a vague conclusion because such a beam could have infinite
number of combinations of rotational inertias at its ends; therefore, a refined model
updating step is to be performed to determine with an acceptable tolerance, the
values of such parameters. As an example, nine rotational-inertia and nine spring-
stiffness combinations are going to be used leaving aside the simply-supported and
clamped-clamped cases, since these are special cases of the spring-supported beam.
The solution space is then composed of eighteen candidates just for this example.
These possible models are going to be correlated with the simulated “experimental”
case whose model is supposedly unknown. The graphical results are shown in Fig. 6.3,
where it is clear that the unknown case (red-colored) is significantly off from any
spring-supported model case (dark-green-colored). This will be evident in the bar
64
plot of the correlation values. On the other hand, when comparing the rotational
mass model cases (blue-colored), a much closer relationship is observed. In fact, one
of the cases gets really close to the experimental case. For a refined result, the analyst
could just add more case combinations to the solution space to get an almost perfect
match.
−1
−0.5
0
0.5
1
Combination 1
(a)
−1
−0.5
0
0.5
1
Combination 2
(b)
−1
−0.5
0
0.5
1
Combination 3
(c)
−1
−0.5
0
0.5
1
Combination 4
(d)
−1
−0.5
0
0.5
1
Combination 5
(e)
−1
−0.5
0
0.5
1
Combination 6
(f)
−1
−0.5
0
0.5
1
Combination 7
(g)
−1
−0.5
0
0.5
1
Combination 8
(h)
−1
−0.5
0
0.5
1
Combination 9
(i)
Figure 6.3. Correlation comparison of spring-supported androtational-mass parameter combinations.
A good preliminary conclusion can be obtained from observation. Indeed, from a
quick look to these, it is clear that for cases (e) and (h), very close correlations between
65
the experimental (red) and rotational mass (blue) models are occurring. However, our
aim is to determine which parameters are causing this correlation; therefore, further
study shall be performed to obtain such values. The algorithm carries out compar-
isons between the experimental model and all the models included in the solution
space, giving one correlation index for each comparison. Each mode shape vector
of the solution space has two additional cells containing the two boundary condition
parameters which can be easily extracted after determining the highest correlation
index. These results are shown in Fig. 6.4(a) and 6.4(b); the first, corresponding to
the spring-supported beam cases; and the second one, to the rotational-inertia cases.
4.3/ 4.3 4.3/ 2.6 4.3/0.43 2.6/ 2.6 2.6/ 1.5 2.6/0.26 0.43/0.43 0.43/0.26 0.43/0.0430
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Parameter Combinations (spring stiffnesses), x 1e6 [N−m/rad],
Cor
rela
tion
Val
ues
(a) Spring-supported
17.2/ 17.2 17.2/ 10.3 17.2/ 3.44 58.2/ 58.2 58.2/ 34.9 58.2/ 11.6 208/ 208 208/ 125 208/ 41.60
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Parameter Combinations (rotational inertias), [N−m2/rad]
Cor
rela
tion
Val
ues
(b) SS w/ rotational masses
Figure 6.4. MAC-based correlation results.
Figure 6.4(b) confirms the preliminary visual conclusion made from Fig. 6.3: that
is to say that combinations 5 and 8 are the closest-related models to the experimental
one because the lines corresponding to the model and the simulation are mostly
66
coincident in these two plots. The final result of this example study is the output
of the model parameters in the command window of MATLAB, which is shown Fig.
6.5.
Figure 6.5. MATLAB command window results.
The obtained parameters are I1 = 52.712 Nm2/rad and I2 = 31.627 Nm2/rad.
The exact values used to simulate the experimental case were I1 = 58.213 Nm2/rad
and I2 = 34.982 Nm2/rad, which yield errors of 9.44% for I1, and 6.1% for I2 and
an overall correlation parameter of 0.9987 which is acceptable to adopt this model
combination as the correct boundary conditions for this experimental setup.
PART III: EXPERIMENTAL VALIDATION
67
7. EXPERIMENTAL SETUP
7.1 Introduction
To experimentally validate the proposed method, a special setup was designed and
constructed. For designing this structure, standard mechanical design procedures
were employed. Not all steps performed will be explained in detail as this is not the
scope of this thesis. The steps [82] include:
1. Conceptualization
2. Research
3. Feasibility assessment
4. Establishing the design requirements
5. Preliminary design
6. Detailed design
7. Production planning and tool design
8. Production
7.2 Conceptualization & Design Parameters
Some important requirements and concepts had to be considered in this structure,
the most relevant were:
1. The dominant behavior of the structure should be similar to those of a beam.
Therefore, a slender and long configuration had to be selected. This condi-
tion derives from the premise that a simple structure was to be used for the
experimental validation.
68
2. The structure should allow both vehicular and pedestrian transit for lab anal-
ysis; therefore it needed to be wide enough to allow for such traffic.
3. The structure should also allow complete and fast assembly-dissasembly of all
its components.
4. The structure had to have the capability to replicate many types of boundary
conditions; namely, pinned, fixed, and even spring-supported configurations in
a fast and effective way.
5. To contribute to damage detection projects which are an important field studied
by the IISL group on a regular basis, the structure should be constructed so
that damage conditions might be simulated for future studies.
6. For realistic modal testing, a first natural frequency around 5 Hz was set as a
requirement.
7.3 Research
To comply with the requirements listed in Sec. 7.2, existing laboratory structures
in similar labs were considered as a reference. Unfortunately, the six conditions
described above made this search a difficult task. Nevertheless, a good approximation
for the needed setup was found in the Smart Infrastructure Management Laboratory
at the University of Florida, which is shown in Fig. 7.1 and whose generous sharing
of information was important in the completion of the present work. This structure
however, did not meet all the requirements because it is a truss structure and its
behavior as a beam was very limited. An additional two structures were considered
but unfortunately neither met all the requirements.
69
Figure 7.1. Scaled truss bridge for SHM. (Courtesy of Dr. JenniferA. Rice, University of Florida at Gainesville).
7.4 Design and Construction
After having discussed many of the above-listed features with fellow members of
the IISL group, a conceptual design was obtained. This design is shown in Fig. 7.2.
The design consists of four 3-section beams connected by endplates with bolts and
nuts; two boundary assemblies, each consisting of a cylindrical shaft between two
pillow block bearings with a pivoting plate connecting the beams with the pin; and
two pseudo-stiffening plates such that all four beams of the bridge have synchronized
vibrations. These plates are called pseudo-stiffening because they don’t provide any
significant structural addition to the behavior of the bridge as their purpose is only
to enforce synchronized vibration.
As shown in the figure, most of the features are included in the conceptual design:
it is a structure that can be modeled as a beam because it does not have any addi-
tional structural elements that would make it behave otherwise; it has discontinuities
70
Figure 7.2. Scaled bridge. Conceptual design.
in each beam which could be considered as damage for further studies; and it has
different boundary condition configuration features.
A finite element model using SAP2000 was developed to test many commercial
steel sections and determine the most suitable one for obtaining the desired first nat-
ural frequency. From this analysis, it was determined that a HSS 3” × 2” × 3/16”
rectangular hollow structural section was the most appropriate. It must be noted
that this analysis was made considering uniform continuous beams and no rotational
masses at the ends. The frequency obtained from the model was around 7 Hz.
Once the sections were selected, the next step in the design process was to develop
detailed drawings. A set of the most relevant details of these drawings is included
in the Appendix. The drawings were submitted to the selected machine and welding
shop for construction under the supervision of the author. Like any other design,
71
some problems were present after the initial assembly that needed to be corrected;
the most important was a twisting of the structure once assembled, due to the weld-
ing process. Manual alignment was performed to correct this issue. Nevertheless, the
received work was acceptable according to the desired expectations.
A special connection was designed for changing the boundary conditions. The
structure has the ability to pivot at the ends through a pin and bearings. It was ex-
pected that the pin can be fixed to the piers through end plate connections whereas
the bearings supply almost frictionless rotation.
To replicate damage in the structure, some steel elements can be replaced by a
similar section of different material. For such a case, three additional aluminum
beam-sections were ordered. Some of the desired effects that these sections provide
are discontinuity of structure, change of stiffness and change of distributed loading
profiles. Finally, the last element is a deck or platform which is not shown in the
figure but will allow safe walking and vehicle traffic over the structure. Vehicle traffic
will be limited only to scaled vehicles. The final dimensions of the bridge were:
• Length = 4500 mm
• Width = 1000 mm
• Height = 500 mm
• Beam section length = 1500 mm
• Approximate weight = 250 kg
7.5 Assembly
The designed structure is 100% disassemblable, for setup versatility and storage
capability. The duration of assembly was around one working day (eight hours), but
72
initial adjustments and corrections took this period to almost two weeks. For the
joints between beam-sections, bolts and nuts were used and the tightening procedure
was performed using a pneumatic power tool to have similar torque in each bolt.
Preliminary torquing of the bolts was made manually, but this caused excessive static
deflection at midspan. Some pictures of the assembled structure will be shown below:
Figure 7.3 depicts the detail of the boundary condition plate. The cylindrical shaft
is the pin which rotates freely when the four bolts (marked with a circle) are not
secured. When the bolts are secured (as in the picture), the boundary turns into a
clamped one. Anchoring of the structure to the floor as well as connections between
beams and end plate are partially shown.
Figure 7.3. Detail of mechanism for modifying the boundary condi-tions: bolted = fixed; unbolted = pinned.
An aerial view of the structure is depicted in Fig. 7.4. The bridge has four beams
(from top to bottom of the figure) and each beam has three sections separated by
connection flanges (marked with two rectangles). For damage simulation, these con-
73
nections can be left loose, a section of the beam can be removed without producing
significant structural instability, or a section can even be replaced by an aluminum
section. The versatility of such a configuration is considerable.
Figure 7.4. Aerial view of the structure.
This section has been intended to be a brief description of the steps taken to have
a feasible experimental setup for the project. To provide detailed design calculations
and material selection procedures is out of the scope of this thesis.
74
8. STRUCTURAL IDENTIFICATION AND CORRELATION OF CASE STUDY
8.1 Introduction
Most modeling problems pose a challenge as the results and findings obtained from
a model always have to be later correlated with experimental observations. As stated
in early chapters, no model is perfect, and paraphrasing George E.P. Box: “All mod-
els are wrong, but some are useful.”
The proposed models are not the exception and from the earliest phase of the
project, this fact was anticipated. Trying to model a complex 3D structure with an
Euler-Bernoulli beam model which is essentially a 2D model represents the first, and
ultimately, the main obstacle. Nevertheless, with the appropriate adjustments, such
a model can be very useful to properly characterize the system’s behavior. In this
chapter, three identification processes will be applied; namely, one assuming that the
bridge is behaving like a simply-supported beam; a second one, assuming the most
critical case of an elastically supported beam which is a clamped-clamped case; a
third one, updating the first two by introducing rotational masses and springs at
both ends accordingly.
The identification will be done by the simple Fast Fourer Transform (FFT) tech-
nique, which will be applied to raw data collected from the structure to determine the
natural frequencies by peak picking. In Section 8.4, three examples of model selection
are developed so that the reader fully understands the mechanics of the method. As
repetitive as this section and its sub sections may seem, they are intended to pro-
vide clarity towards the first step of the technique which is proposed to be a good
model selection. After having understood the previous step, an example with real
75
experimental data gathered from the constructed structure is performed in Section
8.5, where previous knowledge of the model is used as starting point. In this section,
parameter estimation and model refinement are explained. At the end of the chapter,
the reader shall fully understand the two steps needed for a correct application of the
proposed methodology.
8.2 Experimental Setup
A general experimental setup was assembled for testing the structure. Sixteen
PCB PIEZOTRONIC piezoelectric, ceramic shear ICP R© accelerometers were used in
conjunction with two S+O ANALYZER data acquisition boxes with eight channels
per box. Accelerometers were attached to the structure with magnetic mounts in
several configurations which are listed in Table 8.1 and schematically described in
Fig 8.1. For the case of Table 8.1, z corresponds to the vertical direction, longitudi-
nal refers to vibration along the beam’s horizontal x axis, and transversal refers to
vibration along the beams transversal y axis. For this thesis, only configurations 2
and 3 will be analyzed to have a case as close as possible to that of a 2D vibrating
beam.
Table 8.1. Experimental configurations.
Configuration No of sensors Location Figure and symbol
1 16 beam 1 8.1(a), red dot
2 16 beam 2 8.1(a), blue squares
3 16 beam 3 8.1(a), green triangles
4 16 beam 4 8.1(a), orange diamonds
5 14 nodal pts 8.1(b), yellow sand-clocks
6 4 flange A 8.1(b), purple stars
7 4 flange B 8.1(b), red stars
76
(a) Longitudinal configurations.
(b) Transversal configurations.
Figure 8.1. Experimental configurations.
Impact hammer testing was performed in each experiment with a sampling fre-
quency of 1024 Hz to obtain useful frequency information up to half that value
according to Nyquist-Shannonn theorem, i.e., 500 Hz. The hammer used was a PCB
modally tunned hammer with a soft tip for low frequency excitation. Technical details
77
of all the equipment used in all conducted experiments are included as appendices
however, the frequency range of such hammer is shown in Fig. 8.2
The standard experiment is described below:
1. Accelerometer mounting to the structure either using magnetic mounts or wax,
depending if the added weight of the magnet is going to induce a parameter
modification or it could be neglected. For this experiment, the added mass of
accelerometer mounts is negligible.
2. Wiring between sensors, hammer and DAQ boxes. The type of accelerometers
used for the experiment work with 10-32 low-noise type coaxial cables.
3. Initial settings in the data acquisition software including number of boxes to
be used, active/inactive channels, type of signal (AC,DC), sampling frequency,
time frame of each acquisition, data format, sensitivity of instruments, and units
of measurement.
4. Impact testing. The structure is hit repeatedly, for around twenty times, leaving
time between hits so that it can get back to zero vibration before the next hit
(for this structure, usually 10-15 seconds) for around 20 times. The more hits,
the better the results due to more averaging of impact response signals
5. After the first impact testing practice, data must be checked to verify its quality,
consistency between hits, response form and whether it is meaningful. Only
after this step, repeated testing can begin.
6. Post processing. With all the data sets properly stored, processing them with
any selected technique can be performed. In this thesis, as mentioned earlier,
the fast Fourier transform is used.
78
Figure 8.2. Hammer response curves (from PCB Piezotronics).
8.3 Algorithm for Modal Identification Fast Fourier Transform
The goal of modal identification is to obtain modal parameters of a structure such
as natural frequencies, mode shapes and damping ratios from experimental data. As
explained in Chapter 2, many techniques are available both in the frequency and
time domains. Natural frequencies were estimated using the fast Fourier transform
of the response acceleration signal of each accelerometer. A MATLAB code was used
with inputs of impact signal and response acceleration data, and outputs of frequency
response function (FRF or TF), real and imaginary TF, input spectra, and coherence
function. Outputs of the code can be conveniently filtered by accelerometer or by
hit number to determine specific information pertaining a desired location of the
structure. The principal technique used in the code was the fast Fourier transform
(FFT), which takes a discrete signal in the time domain and converts it into its discrete
frequency-domain representation. This technique is extremely useful for computing
modal parameters, as it shows resonant frequencies of a response signal as peaks which
79
are easy to identify. It is also popular because most of real life structures cannot be
tested on a shake table or with a shaker to excite the desired frequencies. An example
of the input and output signals for an arbitrary location of the bridge is shown in Fig.
8.3: the top graph is the impact signal, whereas the bottom graph is the acceleration
response.
0 50 100 150 200 250
100
200
300
400
500
600
Time(s)
Fo
rce(
N)
(a) Hammer impact history (input signal).
0 50 100 150 200 250−15
−10
−5
0
5
10
15
Time(s)
Acc
eler
atio
n(m
/s2 )
(b) Acceleration history (output signal).
Figure 8.3. Experimental input/output signals.
This constitutes the first step in the algorithm, so the analyst can make sure
everything is normal in the input and output.
The next step is to choose which hits are good for analysis and which have to be
neglected. As a manual technique, hammer hitting is not free from errors; at first, it
80
would appear to be an easy task, but it usually takes some time before one becomes
experienced with the instrument to obtain the right hit. Figure 8.4 depicts the impact
signal for a complete test, in which it is evident that impact peaks 9, 10 and 11 are
in the same hit, as well as peaks 14 and 15. For error minimization, these have to be
eliminated from the algorithm, to minimize errors.
0 50 100 150 200 2500
100
200
300
400
500
600
700
800
900
1000
Time(s)
Fo
rce(
N)
←1 ←2←3
←4←5
←6
←7 ←8 ←9
←10
←11
←12
←13
←14←15
←16
←17←18
←19
←20
←21 ←22←23
←24
Figure 8.4. Hammer hit peak choosing.
A second check for choosing the right hits is to analyze each hitting interval, i.e.,
the force versus time of each hit within the chosen input window. There must be
only one hit in the window for the algorithm to produce good results. If two or more
impacts are observed, the hit has to be discarded as well. Figure 8.5 shows a plot in
which all the hits within the chosen window have a single peak, the solid blue line is
the window and the superimposed lines are all the hits. As two or more peaks appear,
the hit should not be included in the analysis.
The next checkpoint is to determine if the impact signal is large enough to excite
all the frequencies of interest. A good technique for this verification is to get the
hammer response curve for the experiment, similar to Fig. 8.2. This plot will deter-
mine if the chosen tip is suitable for exciting the modes of interest. Of course, this
is intended when specific frequencies are to be studied. Nevertheless, time, resources
and effort are used in every experiment, and it is not desirable to waste them by not
81
1 2 3 4 5 6 7 8
0
100
200
300
400
500
Time(s)
Acc
eler
atio
n(m
/s2 )
Hit 1Hit 2Hit 3Hit 4Hit 5Hit 6Hit 7Hit 8Hit 9Hit 10Hit 11Hit 12Hit 13Hit 14Hit 15Hit 16Hit 17Hit 18Hit 19Hit 20
Figure 8.5. Number of peaks within an imput window.
paying attention to all the details.
Just like the input signal needs to be thoroughly analyzed, the output signal must
be taken care of as well. One of the most important aspects of impact testing is the
use of a good window for the response (different than the input window described
above). This resource is necessary due to the fact that a lightly-damped structure,
such as the experimental setup, will not cease vibrating within the short time be-
tween hits. Even if this time is extended as much as possible, very small amplitude
vibrations will remain on the system for rather long time. To overcome this prob-
lem, the signal must be forced to decay to zero within a reasonable window width.
The most popular impact weighing window used for this purpose is the exponential
which is used to “force the data to satisfy the periodicity requirements of the Fourier
Transform” [83].
After all these “sanity checks”, one is ready to obtain the plots of interest: the
FRF, and the imaginary transfer function from which the mode shapes and natural
frequencies and damping ratios can be determined.
82
8.3.1 Modal Analysis Results of Experimental Setup
The MATLAB algorithm output is a set of plots for fully understanding the re-
sponse of the system in the frequency domain. Some of these plots will be shown
herein; specially the two most important: FRF and mode shapes. It should be noted
that a preliminary definition of the peak region must be made for the algorithm to
determine the exact frequency at which a mode appears. This is done in the program
setup and is very useful for modal verification. Figure 8.6 shows the most relevant
code outputs:
0 50 100 150 200 250 300 350 400 450 500
−40
−30
−20
−10
0
10
20
30
40
50
60
Frequency(Hz)
dB
(a) Input Response Curve. All sensors
0 5 10 15 20 25 30−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
Frequency(Hz)
Mag
nitu
de
(b) Imaginary part of FRF.
0 5 10 15 20 25 30 35 40 45 50
−60
−40
−20
0
20
40
60
Frequency(Hz)
Mag
nitu
de(d
b)
(c) Transfer Function. All sensors.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
1
ω1=3.6406 Hz
ω3=25.0078 Hz
ω2=13.8125 Hz
Am
plitu
de
x/L
(d) First Three Modeshapes.
Figure 8.6. FFT code outputs.
Figure 8.6(a) shows the input response for the chosen hammer tip. This means
that frequencies up to 400 Hz can be excited by the hitting action. The imaginary
part of the FRF is plotted in Figure 8.6(b). This plot is particularly useful for deter-
mining the mode shapes as the phase and relative amplitude of each peak determines
the corresponding phase and amplitude of each sensor position. The better the res-
83
olution (i.e., the number of sensors), the better the mode shape estimation. The
resolution issue was addressed at the end of Ch. 4.
In an ideal testing methodology, the complete FRF matrix could be obtained
by hitting every sensor location and acquiring and post-process the generated data.
Unfortunately, the testing procedure has to be planned considering the time factor
and performing n hammer tests would result in a significant consumption of time (n
is the number of sensors), this is why only one row or column of the FRF matrix was
obtained which provided fairly acceptable mode shape results. Figure 8.6(c) shows the
FRF for all the accelerometers where the first four peaks can be clearly distinguished
and finally, Figure 8.6(d) depicts the mode shape estimation for the first three modes
which were later correlated using the proposed method.
8.4 Correlation to a Simply-Supported Beam
The correlation process has been designed to be fully computational. That is to
say that an experimental model should be compared to a group of candidate models
automatically to determine the closest relationship which is the updated model. To
show the methodology in this thesis, the correlation study will be shown individually
for three example cases. In the first case, the model space consists of a unique element
corresponding to the simply-supported beam: single mode shape vectors along with
stacked vectors are plotted on the same axes, as explained in Chapter 4; and finally,
all vectors: single modes and stacked modes are computationally correlated. For
demonstration purposes, a group of superimposed plots will be presented in which all
three modes are compared separately as well as stacked (similar to Fig. 6.1). The
first three normalized mode shape vectors, both in amplitude and in distance of a
simply-supported beam with a resolution of sixteen accelerometers (plus two nodes
at the ends) are shown in Fig. 8.7, 8.8 and 8.9. Each figure depicts a mode shape
column vector with its correspondent graphical representation so the reader can better
84
understand what a mode shape vector is and looks like. It must be kept in mind that
these are the analytical model mode shapes, to which the experimental are going to
be compared. This explanation will be made only for the simply-supported case as
the procedure becomes repetitive for further models.
ψ1 =
00.00600.02970.05220.07250.08980.10350.11290.11780.11780.11290.10350.08980.07250.05220.02970.0060
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
x/L
AM
PLI
TU
DE
SS BEAM − 1ST MODESHAPE
Figure 8.7. First mode shape vector along with its graphical repre-sentation. SS model.
These three mode shape vectors will be correlated with the corresponding three
experimental mode shapes with the addition of two stacked mode shape vectors.
The first stacked mode shape vector and its graphical representation is shown in
Figure 8.10. Due to space limitations, the second stacked vector which consists of
54 row-elements will not be displayed, although its graphical representation will be
presented in the results section. After this preliminary step, the procedure becomes
straightforward: five experimental vectors are going to be compared graphically and
iteratively against five analytical vectors to determine the degree of linear relation
between them. The procedure, as described in Sec. 4.1, will be applied herein for this
model case. Two plots will be presented: a first, corresponding to the mode shape
comparison and a second, one corresponding to the correlation matrix.
85
ψ1 =
00.01210.05760.09370.11460.11700.10040.06760.0238−0.0238−0.0676−0.1004−0.1170−0.1146−0.0937−0.0576−0.0121
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
−1
−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
1SS BEAM − 2ND MODESHAPE
AM
PLI
TU
DE
x/L
Figure 8.8. Second mode shape vector along with its graphical rep-resentation. SS model.
8.4.1 Graphic Results for the Simply-Supported Case
Figure 8.11 shows the graphic relation between five experimental vectors versus five
analytic vectors. A close relationship can be observed for the first two modes as the
red and black lines are almost collinear; however, for the third mode and subsequent
sub-plots, this trend begins to disappear and significant discrepancies between lines
are observed.
8.4.2 Correlation Results for the Simply-Supported Case
The correlation matrix is presented graphically in the form of bars. As more mode
shapes are added to the result, more decreasing trend is observed. Figure 8.12 shows
these results.
From this graphic, many conclusions can be discussed.
86
ψ1 =
00.01820.08170.11600.10870.0625−0.0060−0.0725−0.1129−0.1129−0.0725−0.00600.06250.10870.11600.08170.0182
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
−1
−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
1SS BEAM − 2ND MODESHAPE
x/L
AM
PLI
TU
DE
Figure 8.9. Third mode shape vector along with its graphical repre-sentation. SS model.
1. An acceptable correlation is observed between mode shapes 1 and 2. This is
expected as first and second mode shapes are quite similar for all cases.
2. When comparing between the third mode shape vectors, a dramatic decrease
of correlation index is obtained. This is likely to yield a solid conclusion of
incorrect model.
3. The overall correlation index, which can be obtained by simply taking the last
value of the correlation vector as it shows the linear relationship of all the
stacked mode shapes is CI = 0.3952, which is rather low for concluding a
strong relationship.
4. The correlation vector is presented below:
87
ψ4 =
00.00600.02970.05220.07250.08980.10350.11290.11780.11780.11290.10350.08980.07250.05220.02970.0060
00
0.01210.05760.09370.11460.11700.10040.06760.0238−0.0238−0.0676−0.1004−0.1170−0.1146−0.0937−0.0576−0.0121
0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2−1
−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
1
x/L
AM
PLI
TU
DE
SS BEAM − 1ST AND 2ND MODESHAPES STACKED
Figure 8.10. Vector of two stacked mode shapes along with its graph-ical representation. SS model.
CM =
0.9963
0.9534
0.0117
0.9738
0.3952
.
88
0
0.5
1
mo
de1
−1
0
1
mo
de2
−1
0
1
mo
de3
−1
0
1
stac
k1
−101
stac
k2
Simply Supported Experimental
Figure 8.11. Simply-supported vs. experimental mode shapes.
Figure 8.12. Correlation results. Simply-supported model.
8.4.3 Correlation to a Clamped-Clamped Beam
A similar approach to that of the previous section was applied, but using the model
of a clamped-clamped beam instead. The reader should recall that the purpose of
this algorithm is to look for the higest correlation between the experimental mode
89
shape and the chosen mathematical ones. For brevity, the mode shape vectors are
not presented explicitly, instead, only the three first graphical representations of the
mode shapes are shown in Fig. 8.13.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1CC BEAM − 1ST MODESHAPE
AM
PLI
TU
DE
x/L
(a) 1st Modeshape.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1−1
−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
1CC BEAM − 2ND MODESHAPE
AM
PLI
TU
DE
x/L
(b) 2nd Modeshape.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1−1
−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
1CC BEAM − 3RD MODESHAPE
AM
PLI
TU
DE
x/L
(c) 3rd Modeshape.
Figure 8.13. Modeshapes of a clamped-clamped beam.
Although these look rather similar to those of the simply-supported case, the reader
should note a slight difference in the slope at the boundary conditions. Therefore, a
second check is to be performed, using the same experimental mode shape vectors.
90
After performing the correlation check for a second time, the results are shown in
Fig. 8.14. The conclusions for this graphic are:
0
0.5
1m
od
e1
−1
0
1
mo
de2
−1
0
1
mo
de3
−1
0
1
stac
k1
−1
0
1
stac
k2
Clamped−Clamped Experimental
Figure 8.14. Clamped-clamped vs. experimental mode shapes.
Vec 1 Vec 2 Vec 3 Vec 4 Vec 50
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Cor
rela
tion
Val
ues
Figure 8.15. Correlation results. Clamped-clamped model.
1. Again, high correlation values appear for mode shape vectors 1 and 2.
91
2. An even lower correlation value than that of the simply-supported case is ob-
served in mode shape 3.
3. The overall correlation index, is CI = 0.4598, a small improvement from the
last case, but still too low for a conclusive judgment.
4. The correlation matrix is presented below:
CM =
0.9808
0.9794
0.0018
0.9763
0.4598
.
8.4.4 Correlation to a Beam with Rotational Masses at Both Ends
The third preliminary check is made with a beam with rotational masses at both
ends. It is assumed that the mathematical model matches with enough proximity
to the real inertia of the bridge’s ends. However, this is not going to be always the
case and that is expected. The method would have the capability to identify even a
difference like that. When a computational run of this algorithm is performed to an
unknown structure, it would have to be able to identify its physical parameters such
as inertial forces. A beam with rotational masses on both ends was depicted in Fig.
5.5. The three individual mode shapes are shown in Fig. 8.16.
When correlating these modes with those of the experimental model, the results of
Fig. 8.17 are obtained. For the case of the rotational mass model, a much more clear
coincidence is observed between the experimental and model curves. Nevertheless,
this does not mean that the chosen model is correct. This point has to remain very
clear for the reader. The proposed methodology only offers a procedure for choosing a
92
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
(a) 1st Modeshape.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1−0.5
0
0.5
1
(b) 2nd Modeshape.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1−0.2
0
0.2
0.4
0.6
0.8
1
1.2
(c) 3rd Modeshape.
Figure 8.16. First three mode shapes of a beam with asymmetricalrotational masses.
model close enough so that the system’s behavior can be predicted with such model,
but it does not result in a model confirmation technique.
The observations of this third correlation check are summarized in the following
list:
1. This is the best result of all three correlation tests.
93
0
0.5
1
mo
de1
−1
0
1
mo
de2
−1
0
1
mo
de3
−1
0
1
stac
k1
−1
0
1
stac
k2
Figure 8.17. Rotational mass vs. experimental mode shapes.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Cor
rela
tion
Val
ues
Figure 8.18. Correlation results. Rotational mass model.
2. High correlation values appear in mode shapes 1 and 3. A slightly lower corre-
lation value appears in mode shape 2. This can be obesrved in Fig. 8.17 where
curves corresponding to the 2nd mode shape have a small deviation.
3. The overall correlation index, is CI = 0.8570, which is significantly higher than
the preceding cases. Notice that for a total confirmation of correlation, more
modes can be used in the study and no experimental model will have values
close to unity due to external factors and uncertainties.
94
4. The correlation vector is presented below. However, should be noted that the
main difference is established by the high correlation of the third mode shape
in contrast to the two preceding cases where such value was close to zero.
CM =
0.9920
0.6389
0.9346
0.8267
0.8570
.
8.5 Correlation Algorithm to Determine Model Parameters
The previous three sections were devoted to explaining and demonstrating the vi-
sual and general characteristics of the proposed method. A broad model definition
was made based on such observations. However, as explained in Sec. 6.4, this is
a vague conclusion and does not offer the analyst a solid relationship between the
general model and the unknown boundary conditions. To overcome this problem a
similar procedure as the one developed in the cited section is done herein with the
only difference that instead of using a mode shape set generated with a simulation,
a real experimental mode shape set of unknown boundary conditions is considered.
Such mode shape set is shown in Fig. 8.6(d). A similar comparison to the eighteen
cases utilized in Sec. 6.4 is shown in Fig. 8.19. The solid red line corresponds to the
experimental mode shape obtained from hammer testing of the structure, the dashed
blue and green lines correspond to the spring-supported and rotational-inertia mod-
els, respectively. It is evident that closer correlations of the experimental mode shape
with the latter occur, whereas poor correlations with the former, especially in the
third mode shape is observed.
Similarly, as in Sec. 6.4, the second part of the analysis is to study the correlation
index for each inertia combination which was presented in the form of a bar plot. At
95
−1
−0.5
0
0.5
1
Combination 1
(a)
−1
−0.5
0
0.5
1
Combination 2
(b)
−1
−0.5
0
0.5
1
Combination 3
(c)
−1
−0.5
0
0.5
1
Combination 4
(d)
−1
−0.5
0
0.5
1
Combination 5
(e)
−1
−0.5
0
0.5
1
Combination 6
(f)
−1
−0.5
0
0.5
1
Combination 7
(g)
−1
−0.5
0
0.5
1
Combination 8
(h)
−1
−0.5
0
0.5
1
Combination 9
(i)
Figure 8.19. Correlation comparison of nine parameter combinations.
this point it is necessary to point out that this technique is intended to facilitate the
model updating process but it requires that the analyst also apply good engineering
judgment. A certain amount of knowledge of the structure is required beforehand, for
instance the general type of boundaries at which the structure is subjected to (e.g.,
fixed, free, pinned or some combination of these); the general type of loading (e.g.,
concentrated load, distributed or profiled loading, or some combination of these); and
the type of service the structure is intended for (e.g., heavy or light traffic, pedestrian
traffic, or some combination of these). A good check of the validity of the chosen
96
parameters is to compare the natural frequencies with those of the chosen model. If
such frequencies are comparable within a range between 10% and 20% (for at least
the three first natural frequencies), it could be said that reasonable close model for
an unknown system.
The bar plot of the correlation indexes for both rotational inertia and spring-
supported model cases are shown in Fig. 8.20. The final results are:
• Simply supported structure with rotational masses on both ends.
• Parameter values are: I1 = 52.712 Nm2/rad, and I2 = 52.712 Nm2/rad which
produce a correlation index of 0.937.
• The first five natural frequencies compared to those corresponding to the model
are listed in Table 8.2
Table 8.2. Final frequency comparison (in Hertz).
Analytical FEM Exp.
5.7071 5.8918 3.6328
13.3434 14.4385 13.8516
22.1749 25.8347 27.8750
49.8844 50.6847 47.8203
93.2855 93.6796 94.2180
8.6 Discussion
Even though choosing the boundary condition combination case with the highest
correlation may prove to be enough for selecting a model, this conclusion may not be
final as other parameter combinations may present a high correlation value. Consider
97
4.297/ 4.297 4.297/ 2.578 4.297/0.4297 2.578/ 2.578 2.578/ 1.547 2.578/0.2578 0.4297/0.4297 0.4297/0.2578 0.4297/0.042970
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Parameter Combinations (spring stiffnesses), x 1e6 [N−m/rad],
Cor
rela
tion
Val
ues
(a) Simply supported with rotational springs
17.19/ 17.19 17.19/ 10.31 17.19/ 3.438 52.71/ 52.71 52.71/ 31.63 52.71/ 10.54 208/ 208 208/ 124.8 208/ 41.60
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Parameter Combinations (rotational inertias), [N−m2/rad]
Cor
rela
tion
Val
ues
(b) Simply Supported with rotational masses
Figure 8.20. MAC-based correlation results.
for example a case in which two correlation indexes are very close in value. Which
one is correct and could a threshold be defined to evaluate this quantitatively? This
constitutes an important design question. Certainly one could choose to use more
mode shape vectors. Indeed, when using only the first mode shape vector, almost all
the indexes are approximately equal to one. When the second mode shape is added,
some indexes reduce considerably but still several are close to one. If the third mode
shape is added, a bar plot similar to Fig. 8.20 can be observed. In this case, the
yellow and maroon bars are pretty close. By adding a fourth mode shape vector, it is
98
expected that the leading trend of a parameter combination is maintained, confirming
such as the best match.
Low correlation values as those depicted in Fig. 8.20(a) are a clear indication of
insufficient model space size. A poor correlation index average value is signal of the
wrong model selection. Moreover, if by adding more mode shape vectors the trend
does not change, a different model must be selected and added to the correlation
algorithm.
An alternative to the previously mentioned problem is the fact that each modal
contribution’s weight is different in the total response of the system. Indeed, the first
mode has a bigger contribution than the second mode, which has a bigger contribution
than the third mode, and so on. From this standpoint, one could think about applying
a weighting function to each mode so that the stacked mode shape vector varies by
sections instead of in an overall way. This approach could represent a much more
accurate correlation algorithm.
Another good alternative is to modify the resolution of the experiment in such
a way that a variable resolution is obtained. In sections of the structure where the
mode shape behavior is fully understood and predicted, fewer sensors can be installed
in exchange of zones of the structure in which more understanding of the behavior is
needed. This technique could be linked to the weighting approach described above in
the sense that a higher number of sensors can be used when trying to identify higher
modes.
8.7 Implementation of the Methodology
The correlation technique developed in this thesis is an intermediate step between
experimental data and updated model. For full implementation of the method in a
real structure, a diagram with all the suggested steps is presented in Fig. 8.21, this
diagram is a summary of what an analyst must do to implement this technique in
any structure.
99
Figure 8.21. Correlation method. Implementation
100
9. CONCLUSIONS
A novel methodology for establishing an accurate model for predicting the behavior
of the vibrations in a beam has been developed, in which a linear correlation equa-
tion is applied to compare the degree of linearity between an experimentally obtained
mode shape vector and a group of suitable analytical mode shape vectors extracted
from selected model candidates (model space). The procedure is performed iteratively
comparing one by one until the highest correlation index is achieved.
After performing the proposed methodology, if deviations are still present in the
model, classical model updating techniques can be further applied to tweak the re-
sponse to a most accurate state. However, the correlation technique should be enough.
Finding natural-frequency correlations may not yield the most accurate model as
mode shape correlations. However, a good extra check for the analyst is to compare
mode shapes in a first step; and natural frequencies in a second step to approximate
the model as much as possible to the real structure.
Almost all the classical models along with the non-classical combination cases
studied give a high correlation for the first mode shape and fairly high values for the
second mode shape. When analyzing from the third mode shape onwards, is when
correlation study can be clearly distinguished. Therefore, a good result is obtained
when constructing a vector of at least the three first mode shapes stacked together.
Correlation cannot be studied when comparing individual mode shapes as many of
these are similar from one model to another.
101
An updated model in which the manipulable parameters are those corresponding
to the boundary conditions instead of structural or material parameters, is a much
more realistic model to work with. The main source of model discrepancies should be
looked for in the most unknown aspect of a structural system which precisely are its
boundary conditions. This also holds for model variations over lifespan, as boundary
conditions are the most likely aspect of a system to change due to use and wear.
The proposed methodology can be used as a supplementary or complementary
technique. The former, for cases where computational costs and time requirements
are not a big issue; and the latter, where high accuracy of the updated model is
required. Nevertheless, once a model has been narrowed enough with the correlation
method, computational costs can drop significantly if compared to applying the tra-
ditional model updating techniques for the whole process.
Model spaces can be constructed either from analytical equations, or finite element
models; the method does not require a closed-form solution of any model equation,
but only a mode shape vector, which makes the methodology much more user-friendly
and straightforward.
Deviations of the observed natural frequencies from those of the predicted by the
updated model are expected for higher modes. Indeed, Table 8.2 shows good concor-
dance for the first three modes and an increasing deviation for higher modes due to
noise implications and system uncertainties.
The most visible uncertainty of the experiment was welding torsion which was
not included in the model. Probably this is the reason of some of the deviations
described above. This particular uncertainty is likely to be the cause of having the
same rotational inertia value in both sides when these values are not symmetric in
reality.
102
10. RECOMMENDATIONS AND FUTURE DIRECTIONS
1. Many features could be enhanced for future testing. One of the main issues
encountered is the torsion of the beam-sections due to welding. It is well known
that any welding process will inflict residual stresses due to excessive heat pro-
duced. It is recommended a close monitoring of the manufacturing procedure
to minimize such a problem. An industrial procedure known as matrix man-
ufacturing may be used where a pattern is set up beforehand to constrain the
elements that are going to be welded so the stresses won’t twist them in the
process.
2. For experiments with a large amount of accelerometers (> 5), it is highly rec-
ommended the use of wire identifiers. The task of identifying cables among
many similar-looking ones can be very time consuming.
3. If finite element modeling is selected as the source of the elements of the model
space, careful attention must be paid to filter computational modes.
4. It is not necessary to have a big amount of elements in the model space. The
experimental comparison of Ch. 8 was made with six cases for one of the bound-
ary conditions, and fractions of 1, 0.6, 0.4, 0.2 and 0.1 of that value for the other
boundary condition. This set of combinations yielded 36 elements in the model
space. It is recommended to start looking for the extreme cases (2 elements),
and then include a couple in between. Once the model has been narrowed, more
elements can be added as needed.
5. A logic future direction of this research project is to explore simple structures
different than a beam; say, a one or two story structure.
103
6. The present thesis used only mode shape vectors as the comparison elements.
Natural frequencies were not included in the analysis but only as posterior check
points. It is recommended to include natural frequencies as a weighting element
in the methodology for further accuracy of the method.
7. The structure used in this project has the capability of replicating many types
of boundary conditions. Though, only a simply-supported configuration with
rotational masses at both ends was used. Another logic future step is to explore
different boundary conditions in the same structure and to apply the method-
ology in these.
8. Another possible approach for dealing with the problem of more than one high
correlation index would be to explore the possibility of applying a weighting
factor to each mode, proportional to the contribution of such mode to the total
response of the structure.
LIST OF REFERENCES
104
LIST OF REFERENCES
[1] Wikipedia contributors, “Structural Health Monitoring,” Wikipedia,The Free Encyclopedia, http://en.wikipedia.org/w/index.php?title=Structural\_health\_monitoring\&oldid=612319780 (accessed July 23,2014).
[2] Wikipedia contributors, “Modal Analysis,” Wikipedia, The Free Encyclo-pedia, http://en.wikipedia.org/w/index.php?title=Modal\_analysis\&oldid=618123843 (accessed July 23, 2014).
[3] Wikipedia contributors, “Finite Element Updating,” Wikipedia, The FreeEncyclopedia, http://en.wikipedia.org/w/index.php?title=Finite\_element\_updating\&oldid=594129961 (accessed July 23, 2014).
[4] Caicedo, J. M. “Two Structural Health Monitoring Strategies Based on GlobalAcceleration Responses: Development, Implementation, and Verification.”MSc. Thesis, Washington University at Saint Louis, 2001. Department of CivilEngineering, 2001.
[5] Meirovitch, L. “Fundamentals of Vibrations”. International Edition, McGraw-Hill, 2004.
[6] Karnovsky, I. A., O. I. Lebed. “Free Vibrations of Beams and Frames”. NewYork: McGraw-Hill, 2004.
[7] Gorman, D.J. “Free Vibration Analysis of Beams and Shafts”. New York:Wiley, 1975.
[8] Blevins, R. D., and R. Plunkett. “Formulas for Natural Frequency and ModeShape.” Journal of Applied Mechanics 47 (1980): 461.
[9] Hibbeler, R. C. “Free Vibration of a Beam Supported by UnsymmetricalSpring-Hinges.” Journal of Applied Mechanics 42, no. 2 (1975): 501-502.
[10] Hibbeler, R. C. “Erratum: Free Vibration of a Beam Supported by Unsymmet-rical Spring Hinges. (Journal of Applied Mechanics, 1975, 42, pp. 501 502).”Journal of Applied Mechanics 48, no. 2 (1981): 449-449.
[11] Goel, R. P. “Free Vibrations of a Beam-Mass System with Elastically Re-strained Ends.” Journal of Sound and Vibration 47, no. 1 (1976): 9-14.
[12] Rao, C. K. and S. Mirza. “A Note on Vibrations of Generally RestrainedBeams.” Journal of Sound and Vibration 130, no. 3 (1989): 453-465.
[13] Chun, K-R. “Free Vibration of a Beam With one End Spring-Hinged and theother Free.” Journal of Applied Mechanics 39, no. 4 (1972): 1154-1155.
105
[14] Lv, B., W-L. Li, J. Dai, H-J. Zhou, F-X. Guo, and Z-B. Gao. “Vibration Analy-sis of Beams with Arbitrary Elastic Boundary Conditions.” Applied Mechanicsand Materials 66 (2011): 1325-1329.
[15] Kang, K-H., and K-J. Kim. “Modal Properties of Beams and Plates on Re-silient Supports with Rotational and Translational Complex Stiffness.” Journalof Sound and Vibration 190, no. 2 (1996): 207-220.
[16] Register, A. H. “A note on the vibrations of generally restrained, end-loadedbeams.” Journal of Sound and Vibration 172, no. 4 (1994): 561-571.
[17] Yen, T. C., and S. Kao. “Vibration of Beam Mass System with Time-dependent Boundary Conditions.” ASME Journal of Applied Mechanics 26(1959): 353-356.
[18] Laura, P. A., J. L. Pombo, and E. A. Susemihl. “A Note on the Vibrations ofa Clamped-Free Beam With a Mass at the Free End.” Journal of Sound andVibration 37, no. 2 (1974): 161-168.
[19] Maurizi, M. J., R. E. Rossi, and J. A. Reyes. “Vibration Frequencies for aUniform Beam with One End Spring-Hinged and Subjected to a TranslationalRestraint at the Other End.” Journal of Sound and Vibration 48, no. 4 (1976):565-568.
[20] Maurizi, M. J., D. V. Bambill de Rossit, and P. A. Laura. “Free and ForcedVibrations of Beams Elastically Restrained Against Translation and Rotationat the Ends.” Journal of Sound and Vibration 120, no. 3 (1988): 626-630.
[21] Maurizi, M. J., R. E. Rossi, and J. A. Reyes. “Comments on A Note of Gen-erally Restrained Beams.” Journal of Sound and Vibration 147, no. 1 (1991):167-171.
[22] Greif, R., and S. C. Mittendorf. “Structural Vibrations and Fourier Series.”Journal of Sound and Vibration 48, no. 1 (1976): 113-122.
[23] Li, W. L. “Free Vibrations of Beams with General Boundary Conditions.”Journal of Sound and Vibration 237, no.4 (2000): 709-725.
[24] Wang, J-S., and C-C. Lin. “Dynamic Analysis of Generally Supported BeamsUsing Fourier Series.” Journal of Sound and Vibration 196, no. 3 (1996): 285-293.
[25] Yayli, M. O., M. Aras, and S. Aksoy. “An Efficient Analytical Method for Vi-bration Analysis of a Beam on Elastic Foundation with Elastically RestrainedEnds.” Shock and Vibration, vol. 2014, Article ID 159213, 7 pages, 2014.doi:10.1155/2014/159213
[26] Lin, Y. K. “Free Vibrations of a Continuous Beam on Elastic Supports.” In-ternational Journal of Mechanical Sciences 4, no. 5 (1962): 409-423.
[27] Lai, H-Y., J-C. Hsu, and C-K. Chen. “An Innovative Eigenvalue ProblemSolver for free Vibration of Euler-Bernoulli Beam by using the Adomian De-composition Method.” Computers & Mathematics with Applications 56, no.12 (2008): 3204-3220.
106
[28] Mao, Q. “Free Vibration Analysis of Elastically Connected Multiple-Beamsby Using the Adomian modified Decomposition Method.” Journal of Soundand Vibration 331, no. 11 (2012): 2532-2542.
[29] Failla, G., and A. Santini. “A solution Method for Euler-Bernoulli VibratingDiscontinuous Beams.” Mechanics Research Communications 35, no. 8 (2008):517-529.
[30] Bilello, C., and L. A. Bergman. “Vibration of Damaged Beams Under a MovingMass: Theory and Experimental Validation.” Journal of Sound and Vibration274, no. 3 (2004): 567-582.
[31] Garcıa-Garino C., A. E. Mirasso, M. A. Storti, and M. E. Tornello. “Vibra-ciones de Vigas y Porticos Ante la Presencia de Fisuras. Implementacion delProblema Inverso.” Mecanica Computacional no. XXXII (2013): 1715-1735.
[32] Wang, J-L., and P-Z. Qiao. “Vibration of Beams with Arbitrary Discontinu-ities and Boundary Conditions.” Journal of Sound and Vibration 308, no. 1(2007): 12-27.
[33] Ratazzi, A. R., D. V. Bambill, and C. A. Rossit. “Free Vibrations of BeamSystem Structures with Elastic Boundary Conditions and an Internal ElasticHinge.” Chinese Journal of Engineering, vol. 2013, Article ID 624658, 10 pages,2013. doi:10.1155/2013/624658
[34] Courant, R. “Variational Methods for the Solution of Problems of Equilibriumand Vibrations.” Bull. Amer. Math. Soc 49, no. 1 (1943): 1-23.
[35] Huebner, K. H., and E. A. Thornton. “The Finite Element Method for Engi-neers, 1982.” Wiley, New York.
[36] Zienkiewicz, O. C., and R. L. Taylor. “The Finite Element Method for Solidand Structural Mechanics”. Butterworth-Heinemann, 2005.
[37] Salgado da Silva Oliveira, J. M. “Analise Modal Experimental Aplicada a umComponente Estrutural Automovel.” (2012). http://hdl.handle.net/1822/22703
[38] Fu, Z-F., and J-M. He. “Modal Analysis”. Butterworth-Heinemann, 2001.
[39] Visser, W. J. “Updating Structural Dynamics Models Using Frequency Re-sponse Data.” PhD Dissertation., University of London, 1992.
[40] Brown, D. L., and R. J. Allemang. “The Modern Era of Experimental ModalAnalysis.” Sound and Vibration 41, no. 1 (2007): 16-33.
[41] Mendes Maia, N. M., and J. M. Montalvao e Silva. “Theoretical and Experi-mental Modal Analysis”. Taunton: Research Studies Press, 1997.
[42] Ewins, David J. “Modal Testing: Theory and Practice”. Vol. 79. Letchworth:Research Studies Press, 1984.
[43] Fu, Z-F, and J-M He. “Modal Analysis”. Butterworth-Heinemann, 2001.
107
[44] Cooley, J. W., and J. W. Tukey. “An Algorithm for the Machine Calculationof Complex Fourier Series.” Mathematics of Computation 19, no. 90 (1965):297-301.
[45] Mendes Maia, N. M. “Modal Identification Methods in the Frequency Do-main.” In Modal Analysis and Testing, pp. 251-264. Springer Netherlands,1999.
[46] Bishop, R. E. D. “An Investigation into the Theory of Resonance Testing.”Philosophical Transactions of the Royal Society of London. Series A, Mathe-matical and Physical Sciences, no. 1055 (1963): 241-80.
[47] Kennedy, C. C. “Use of Vectors in Vibration Measurement and Analysis.”Journal of the Aeronautical Sciences (Journal of the Aeronautical Sciences)14, no. 11 (1947): 603-625.
[48] Dobson, B. J. “Modal Analysis Using Dynamic Stiffness Data.” 3rd Interna-tional Modal Analysis Conference, Orlando, FL. (1985): 279-285
[49] Dobson, B. J. “A Straight-Line Technique for Extracting Modal Propertiesfrom Frequency Response Data.” Mechanical Systems and Signal Processing1, no. 1 (1987): 29-40.
[50] Mendes Maia, N. M. “Extraction of Valid Modal Properties from MeasuredData in Structural Vibrations.” PhD Dissertation, Imperial College London(University of London), 1988.
[51] Gaukroger, D. R., C. W. Skingle, and K.H. Heron. “Numerical Analysis ofVector Response Loci.” Journal of Sound and Vibration 29, no. 3 (1973): 341-353.
[52] Ewins, D. J., and P. T. Gleeson. “A Method for Modal Identification of LightlyDamped Structures.” Journal of Sound and Vibration 84, no. 1 (1982): 57-79.
[53] Brittingham, J. N., E. K. Miller, and J. L. Willows. “Pole Extraction fromReal-Frequency Information.” Proceedings of the IEEE 68, no. 2 (1980): 263-273.
[54] Schmerr, L. W. (1982), “A New Complex Exponential Frequency DomainTechnique for Analysing Dynamic Response Data, Proceedings of the 1st In-ternational Modal Analysis Conference (IMAC I), Orlando, Florida, USA,183-186.
[55] Juang, J-N., and H. Suzuki. “An Eigensystem Realization Algorithm in Fre-quency Domain for Modal Parameter Identification.” Journal of Vibration,Acoustics Stress and Reliability in Design 110, no. 1 (1988): 24-29.
[56] Richardson, M. H., and D.L. Formenti. “Parameter Estimation from Fre-quency Response Measurements Using Rational Fraction Polynomials.” InProceedings of the 1st International Modal Analysis Conference, vol. 1, pp.167-186. 1982.
[57] Fillod, R., G. Lallement, J. Piranda, and J. L. Raynaud. “Global Method ofModal Identification.” In Proceedings of the 3rd IMAC Conference, Orlando,Florida, pp. 1145-1151. 1985.
108
[58] Zhang, L. and H. Kanda. “The Algorithm and Application of a New Multi-Input-Multi-Output Modal Parameter Identification Method.” Shock and Vi-bration Bulletin, (1986): pp. 11-17
[59] Coppolino, R. N. “A Simultaneous Frequency Domain Technique for Estima-tion of Modal Parameters from Measured Data”. No. 811046. SAE TechnicalPaper, 1981.
[60] Link, M., and A. Vollan. “Identification of Structural System Parameters fromDynamic Response Data.” Zeitschrift fur Flugwissenschaften und Weltraum-forschung 2 (1978): 165-174.
[61] Klosterman, A. L. “On the Experimental Determination and Use of ModalRepresentations of Dynamic Characteristics”. Order No. 7216547, Universityof Cincinnati, 1971. http://search.proquest.com/docview/302560639? ac-countid=13360
[62] Leuridan, J., and J. Kundrat. “Advanced Matrix Methods for ExperimentalModal Analysis-A Multi-Matrix Method for Direct Parameter Extraction.” InProceedings, International Modal Analysis Conference (1982): 192-200.
[63] Juang, J-N., and Richard S. Pappa. “An Eigensystem Realization Algorithmfor Modal Parameter Identification and Model Reduction.” Journal of Guid-ance, Control, and Dynamics 8, no. 5 (1985): 620-627.
[64] Spitznogle, F. R., and A. H. Quazi. “Representation and Analysis of Time-Limited Signals Using a Complex Exponential Algorithm.” The Journal of theAcoustical Society of America 47, no. 5A (2005): 1150-1155.
[65] Vold, H., and G. T. Rocklin. “The Numerical Implementation of a Multi-Input Modal Estimation Method for Mini-Computers.” In International ModalAnalysis Conference Proceedings (1982): 542-548.
[66] Ibrahim, S. R., and E. C. Mikulcik. “A Time Domain Modal Vibration Tech-nique.” The Shock & Vibration Bulletin, vol. 46(5) (1976): 187-196.
[67] Gersch, Will. “Estimation of the Autoregressive Parameters of a Mixed Au-toregressive Moving-Average Time Series.” Automatic Control, IEEE Trans-actions on 15, no. 5 (1970): 583-588.
[68] Gersch, W., N. N. Nielsen, and H. Akaike. “Maximum likelihood estimationof structural parameters from random vibration data.” Journal of Sound andVibration 31, no. 3 (1973): 295-308.
[69] Leuridan, J., and H. Void. “A Time Domain Linear Model Estimation Tech-nique for Multiple-Input Modal Analysis.” In Modal Testing and Model Re-finement: Presented by the Winter Annual Meeting of the American Societyof Mechanical Engineers, Boston, Massachusetts, November 13-18, (1983): 51.American Society of Mechanical Engineers, 1983.
[70] Juang, J-N, and M-Q. Phan. “Identification and Control of Mechanical Sys-tems”. Cambridge University Press, 2001.
[71] Allemang, R. J. “The Modal Assurance Criterion-Twenty Years of Use andAbuse.” Sound and Vibration 37, no. 8 (2003): 14-23.
109
[72] Lieven, N. A. J., and D. J. Ewins. “Spatial Correlation of Mode Shapes, theCoordinate Modal Assurance Criterion (COMAC).” In Proceedings of the 6thinternational Modal Analysis Conference, vol. 1 (1988): 690-695.
[73] Heylen, W., and S. Lammens. “FRAC: a Consistent Way of Comparing Fre-quency Response Functions.” In Proceedings of the Conference on Identifica-tion in Engineering Systems, (1996): 48-57.
[74] Fotsch, D., and D. J. Ewins. “Application of MAC in the Frequency Domain.”ROLLS ROYCE PLC-REPORT-PNR (2000).
[75] Heylen, Ward, and Theo Janter. “Extensions of the modal assurance crite-rion.” Journal of Vibration and Acoustics 112, no. 4 (1990): 468-472.
[76] Brechlin, E., K. Bendel, and W. Keiper. “A New Scaled Modal AssuranceCriterion for Eigenmodes Containing Rotational Degrees of Freedom.” In Pro-ceedings of the International Seminar on Modal Analysis, vol. 3, (1999): 1175-1182. Katholieke Universiteit Leuven
[77] Messina, A., I. A. Jones, and E. J. Williams. Proceedings of Conference onIdentication in Engineering Systems, Swansea, U.K., 67/76. “Damage detec-tion and localisation using natural frequency changes”
[78] Messina, A., T. Contursi, and E. J. Williams. “Multiple Damage EvaluationUsing Natural Frequency Changes.” Proceedings of the 15th InternationalModal Analysis Conference 1, (1997): 658-664
[79] Koh, B. and S. J. Dyke. “Structural Health Monitoring for Flexible BridgeStructures using Correlation and Sensitivity of Modal Data. Computers &Structures, (3-4), (2007): 117-130.
[80] Pandey, A. K., and M. Biswas. “Damage Diagnosis of Truss Structures byEstimation of Flexibility Change.” Modal Analysis-The International Journalof Analytical and Experimental Modal Analysis 10, no. 2 (1995): 104-117.
[81] MATLAB Release 2013a, The MathWorks, Inc., Natick, Massachusetts,United States.
[82] Wikipedia contributors, “Engineering Design Process,” Wikipedia, The FreeEncyclopedia, http://en.wikipedia.org/wiki/Engineering\_design\_process
[83] Avitabile, P. “Modal Space-In Our Own Little World.” Experimental Tech-niques 37, no. 2 (2013): 3-5.
[84] Chazi, Eleni. “Identification Methods for Structural Systems.” Course Lec-tures. http://www.vvz.ethz.ch/Vorlesungsverzeichnis/dozentPre.do?semkez=2014S&dozide=10005329&ansicht=2&lang=en Eidgenossische Tech-nische Hochschule Zurich, 2013.
[85] Chopra, Anil K. Dynamics of structures. Vol. 3. New Jersey: Prentice Hall,1995.
VITA
110
VITA
CHRISTIAN E. SILVA
CAREER OBJECTIVE
A Master of Science in Mechanical Engineering with concentration in Structural Dy-
namics & Vibrations, and Control Systems.
EDUCATION
Purdue University Master of Science in Mechanical Engineering. Candidate August
2014. Grade Point Average 3.43 on a 4.00 scale
Escuela Superior Politecnica del Litoral, Ecuador. Mechanical Engineer. February,
2002. Grade Point Average 7.11 on a scale of 10.
WORK EXPERIENCE Ecuasteel S. A., Guayaquil-Ecuador August 2004 -
July 2012
• Industrial machinery construction project management.
• General management tasks for manufacturing department and administrative
areas.
IIASA-CAT, Guayaquil-Ecuador August 2002 - August 2004
• Engine Technical Department Supervisor.
HONORS AND ACTIVITIES
• Ecuadorians at Purdue University Association - President 2014 - Present.
• Secretarıa Nacional de Ciencia, Tecnologıa e Innovacion, Ecuador (National Sec-
retary of Science Tecnology, and Innovation Ecuador) - Universities of Excellence
Scholarship Program. Grantee, Aug 2011.
top related